AINEWS - OpenRouter Discord

OpenRouter ▷ #announcements (1 messages):

Kimi K2, Crashloop, Issue Resolution

• Kimi K2 Thinking Crashloop Crisis Averted: Kimi K2 Thinking experienced issues with both providers due to a crashloop triggered by a specific prompt.
  • The issue has been resolved after collaborative efforts.
• Prompt-Induced Crashloop Plagues Kimi K2: A crashloop, induced by a problematic prompt, caused issues for Kimi K2 across multiple providers.
  • Teams are actively collaborating to pinpoint and squash the gremlin causing the outage.

OpenRouter ▷ #app-showcase (7 messages):

Orchid AI Assistant, Release Date Estimation, The nature of work

• Orchid AI Assistant ETA: 2-48 Months!: The estimated release date for Orchid AI Assistant is projected to be within the next 2 to 48 months.
  • A member reacted to this long and vague estimate with the word “crazy”.
• Contemplating the nature of ‘work’: A member expressed a dislike for “working,” suggesting that AI development aims to address this sentiment.
  • The statement implies a desire to automate or alleviate the burdens associated with traditional labor through AI technologies.

OpenRouter ▷ #general (569 messages🔥🔥🔥):

OpenRouter video support, Polaris Alpha mini model, OpenAI adult content handling, Kimi K2 leaderboard rankings, Gemini 2.5 token usage

• OR may support videos in the future: A user expressed a wish for OpenRouter to support videos and text-to-speech (TTS) functionality, as shared in this tweet.
• Polaris Alpha possibly not a mini model: There is speculation that Polaris Alpha might not be a mini model, contrasting with the approach OpenAI took with GPT-5 as outlined in the GPT-5 System Card.
• OpenAI going adult - impacts OpenRouter: There is a question of how OpenRouter will handle OpenAI allowing adult content for users over 18, and whether users will need to bring their own API keys.
• Gemini 2.5 Flash chews through tokens: A user found that a 24-second, 900x600 video uploaded to Gemini 2.5 Flash consumed over 800k input tokens, contrary to Google’s documentation stating fixed rates of 263 tokens per second for video as mentioned in the token documentation.
• Cerebras Mandatory Reasoning: Users reported issues with the Cerebras model, where disabling reasoning caused errors; documentation confirms reasoning is mandatory.
  • One workaround suggested was to omit the reasoning parameter altogether, after finding that enable should be enabled in the parameters.

OpenRouter ▷ #new-models (2 messages):

“

• No New Models Discussion: There was no discussion about new models in the provided messages.
  • The channel appears to be empty or the messages are not relevant to the topic.
• Absence of Relevant Content: No specific details or links related to model updates or technical discussions were found.
  • The content might be missing or requires more context to generate meaningful summaries.

OpenRouter ▷ #discussion (29 messages🔥):

OpenRouter Model Node on n8n, OR Show Technical Segment, GPT-4 Regression, Chatroom Memory Setting, Automated Capabilities Scanning

• OpenRouter Model Node Inquiries Spark Curiosity: Members inquired whether the OpenRouter model node on n8n was created by the OpenRouter team or by an external entity.
  • Another member suggested including a brief technical segment in the OR Show, such as a screen recording with a short discussion.
• GPT-4 Regression Troubles Users: Users reported a regression from GPT-4, with one noting that they were surprised to see the issue, and another saying Claude found two other discrepancies.
  • The thread included attached images documenting discrepancies between different models on the platform.
• Chatroom ‘Memory’ Setting Misunderstood: A user inquired about the ‘chat history’ setting in the Chatroom, renamed as ‘Memory’, wondering what happened to it since the default value is 8.
  • Another user clarified its location at the bottom, noting it was previously in the top left tab button, and another that they thought that would actually limit the $120/mtok output somehow.
• Automated Capabilities Scanning Proposed: A member suggested implementing some kind of automated capabilities scanning to detect changes in models/providers over time.
  • They linked to an article on Cursor as an example, describing how a basic getWeather tool call could be used to check for functionality changes.
• GPT-5 Excels, Gemini Falls Flat: A user shared their positive experience using GPT-5 for creating schedules with nomenclature and filename structures, while noting their negative experience being locked out of Gemini code assist due to quota issues.
  • They also mentioned needing to use DS3.1 for john-the-ripper help because Kimi refused, praising Meta’s under-the-radar AI projects and linking to a post on X to illustrate their point.

OpenRouter ▷ #announcements (1 messages):

Kimi K2 Thinking, MoonShot AI, Test-time scaling, Agentic performance

• MoonShot AI drops Kimi K2 Thinking Model: MoonShot AI released their new thinking model Kimi K2 Thinking, claiming SOTA on HLE (44.9 %) & BrowseComp (60.2 %).
  • This model executes 200–300 tool calls autonomously, excels in reasoning, agentic search, and coding, and features a 256K context window.
• Kimi K2 Thinking boasts Test-Time Scaling: Kimi K2 Thinking is trained for test-time scaling, interleaving thought and tool use over long sequences for stable, goal-directed reasoning.
  • For the best agentic performance, users are instructed to return reasoning content upstream (reasoning_details field) so the model can see its own thinking steps and maintain coherence across calls, according to OpenRouter docs.

OpenRouter ▷ #app-showcase (5 messages):

image generation failures, cat girl images

• Image Generation Tests are Failing: A member reported that image generation tests are failing with the message “Not working”.
  • The issue may be specific to the UK.
• Cat Girl Image Sparks Approval: A member suggested that a generated image including a cat girl is of high quality.
  • They added the comment “there’s a cat girl in the image, so it’s gotta be good”.

OpenRouter ▷ #general (176 messages🔥🔥):

OpenRouter downtime, Qwen3 Rate Limits, GPT-5 Image Mini Issues, Apple using Google AI for Siri, DeepSeek OCR integration

• OpenRouter has Pumpkin Icon Downtime Drama: Users reported timeout 408 errors across various models, with some unable to check credits, and one user noticed a pumpkin icon.
  • Some users humorously suggested switching to local models or generating tokens in their heads while waiting for a fix.
• Qwen3 Coder Free Model Faces Rate Limit Frustration: Users experienced consistent rate limit errors with the Qwen3 Coder Free model, even after weeks of inactivity, leading to frustration that paid credits are not improving the rate limit.
  • It was clarified that the free model shares rate limits among all users, so some users are unlikely to get a request through, and were recommended trying paid models like glm 4.6/4.5, Kimi K2 or Grok code fast.
• GPT-5 Image Mini Model Magically Stops Image Generation: A user reported that the gpt-5-image-mini model stopped generating images in both the chatroom and API, and the activity page showed minimal image output.
  • It was unclear whether the issue was account-specific or a broader problem with OpenRouter.
• Apple’s Siri to Potentially Embrace Google’s AI: A user shared a Reuters article stating that Apple plans to use a 1.2 trillion-parameter AI model developed by Google to revamp Siri.
  • Discussion was terse, and users pointed to other more important priorities.
• DeepSeek-OCR Desired by Document-Loving Devotees: A user suggested integrating DeepSeek-OCR into OpenRouter, praising its powerful document processing capabilities and OCR performance.
  • The user noted that several others have requested the integration of this model.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussion: There was no discussion about new models in the given messages.
• No candidate topics identified: The provided messages did not contain any specific candidate topics suitable for detailed summaries.

OpenRouter ▷ #discussion (26 messages🔥):

Tiger Data Agent Cookout, Claude Prompt Jailbreak, GPT Model Censorship, OpenAI Codex Update, OpenRouter Chatroom Issues

• Tiger Data Throws Coding Agent Cookout: The Tiger Data team is hosting an agent cookout in Brooklyn, NY on November 13th, from 6-9 pm to build coding agents and chat with their engineering team, with RSVP link here.
• Users Discuss Ways to Circumvent Claude’s Ethical Restrictions: Users discussed prompt jailbreaks for Claude to bypass its “inaccurate ethical concerns”, with one suggesting using GPT 4.5 to create a safe script and then asking Claude to correct it by adding “criminal code”.
  • Another user commented that this is effective because “Claude likes to correct its mistakes”.
• New Desertfox Model pushed to OpenAI Codex: A member mentioned that a new model called desertfox was pushed to OpenAI Codex, linking to the relevant GitHub commit.
• OpenRouter Chatroom Glitches Reported: A user reported that the OpenRouter chatroom was broken, with a link to the specific chat model page.
• GPT-7 Might Beat GTA 6 to Release: A user joked that GPT-7 might be released before GTA 6, citing another delay for the game.

OpenRouter ▷ #announcements (3 messages):

OpenRouter Charts, Activity Grouping, Filtering Options

• OpenRouter Charts Get More granular: Members are excited at the news that OpenRouter charts can now group activity by user and API key.
  • This new feature provides more detailed insights into usage patterns and cost allocation.
• A Week View Requested: Users request the addition of a one-week filtering option to the new OpenRouter charts.
  • This would allow for easier tracking of weekly usage trends, as illustrated in an attached image.

OpenRouter ▷ #app-showcase (2 messages):

fenic, OpenRouter Integration, LLM ETL, AI Workflows

• Fenic Integrates with OpenRouter!: Fenic, a dataframe API and execution engine for AI workflows, now integrates with OpenRouter to enable running mixed-provider pipelines in one session.
  • The integration aims to help scale large batches cleanly and swap models without touching pipeline code, as well as unblock a broader landscape of models for LLM ETL, context engineering, agent memory, and agent tooling.
• Typedef-AI’s fenic Library: Typedef-AI announces the release of their new library, fenic, which provides a dataframe API and execution engine for AI workflows.
  • It aims to streamline tasks such as LLM ETL, context engineering, agent memory, and agent tooling.

OpenRouter ▷ #general (527 messages🔥🔥🔥):

ComfyUI for Free, AMD vs Nvidia for LLMs, Model Context Limits, Deepseek and Roleplay

• Cloud GPU Costs Spark Debate: Members discuss the value of paying for cloud GPU rentals for AI image generation, with one user preferring to stick with slower but free, local options using their AMD cards.
  • One user stated “don’t tell me to rent gpu on cloud i dont think its even worth throwing a cent at ai”.
• Ollama Simplifies Model Testing: A member recommends using Ollama for simple setup and testing of various models, particularly on a desktop after switching to Linux to avoid Windows issues.
  • Another member pointed out that “the setup is pretty simple and they have a library of models to download from”.
• GPT’s accuracy disputed: Users question GPT’s accuracy regarding AMD cards and model performance, citing feedback from stable diffusion communities indicating AMD cards are viable, despite GPT’s claims.
  • One user expressed a lack of trust after GPT incorrectly stated AMD cards were worse for stable diffusion despite community feedback.
• Model Context Woes plague roleplayers: A user complains about models for roleplay quickly running out of context and “forgetting” key details after only 30-40 messages.
  • One user pointed out that the issue might be the default token limit in LM Studio, suggesting that they “need to change that in the model’s config”.
• OCR + Gemini Pro = CYOA Sheet Solution?: Users suggest using Gemini Pro and OCR to extract text from CYOA sheets for AI processing, due to difficulties in parsing images directly.
  • A user stated that Gemini 2.5 pro is pretty good image readers you just need to use it when you need your sheet in text**.

OpenRouter ▷ #discussion (100 messages🔥🔥):

Google's AI Model Dislike, Bedtime Fable Animation Engine, Gemma Models Solve Captchas, Provider Feedback System, Movement Labs Allegations

• Google’s AI Dislike Becomes Courtroom Drama: A user expressed hope that Google will argue in court that its AI model dislikes a specific individual due to their negative online presence, referencing a tweet.
  • The same user also suggested feeding depositions into an animation engine to create bedtime fables, linking to a YouTube video as an example.
• Gemma Models Shine at Captcha Cracking: A user noted that Gemma models are surprisingly effective at solving captchas.
  • Another user proposed implementing a dislike button next to each provider under each model for feedback, along with a general feedback button.
• Like/Dislike Buttons Spark Debate: The idea of like/dislike buttons for providers was discussed, with users debating the merits of binary feedback vs. comment systems.
  • Some argued that comments could become toxic, while others worried about rating systems being gamed or downvote bombed.
• Movement Labs Face Skepticism and “Scam” Accusations: Users expressed skepticism towards Movement Labs, with one user calling it a potential scam.
  • Concerns were raised about their claims of enabling models up to 34.6 trillion parameters and attempts to prove they aren’t a Cerberas wrapper, with users highlighting suspicious behavior and marketing language.
• Continuous Benchmarks: One user suggests implementing continuous benchmarks and transparent uptime metrics for providers.
  • They recommended displaying data from the K2-Vendor-Verifier GitHub repository beneath each provider to show tool calling success rates and performance on benchmarks like GPQA Diamond or MMLU Pro.

OpenRouter ▷ #announcements (3 messages):

OpenRouter charts, Activity grouping, Filtering Options

• OpenRouter Adds More Charts: OpenRouter announced on X the addition of more charts, grouping activity by user and API key.
• Users Request More Filtering Options: A user expressed their appreciation for the new charts and suggested adding a one-week filtering option.

OpenRouter ▷ #app-showcase (11 messages🔥):

Fun Website with API key, Frontend AI, OpenRouter Integration

• Fun Website uses OpenRouter API Key: A member created a fun website that utilizes the OpenRouter API key and allows users to choose their model, tested on Kimi 0905 with groq.
  • The API key is stored locally with a privacy policy, the website is open-sourced at github.com/koikbr/web98.
• Frontend AI tweaks: A member suggested some tweaks to the website’s frontend, such as avoiding blue to purple gradients and emoji-annotated bullet lists, as they resemble AI-generated content.
  • They suggest the risk level listings at huggingface.co/openguardrails/OpenGuardrails-Text-2510 give them the ick, specifically in relation to harm to minors or insulting national symbols.
• Dataframe API fenic Integrates with OpenRouter: A member shared that fenic, a dataframe API and execution engine for AI workflows, now integrates with OpenRouter.
  • This integration enables users to run mixed-provider pipelines, scale large batches, swap models, and unblock a broader landscape of models.

OpenRouter ▷ #general (684 messages🔥🔥🔥):

GLM 4.6, DeepInfra quantization, Sapphira-L3.3-70b-0.1, OpenRouter Presets, OpenRouter embedding models

• Scammer exploits OpenRouter’s bio section: A user flags a non-working website in another member’s bio (ultra.doan) as a scam and a sham, leading to the bio-owner admitting being too lazy to fix it.
  • The user also noted they were simply renewing the domain to keep their brand ✨.
• OpenRouter embeds now flying onto scene: A user asks for docs on embeds, excitedly quoting They fly now, referring to recent addition of embedding models to OpenRouter.
  • Members responded with quips like They float and They helicopter it, celebrating the new feature.
• Amazon’s pricing triggers tears: A user complains about Amazon’s pricing at $12.5/M, lamenting that it’s so high nobody will pick it over Sonnet 4.5.
  • Other members echoed the sentiment, with one stating I don’t need to test it to know that about the model’s quality.
• Token Tariffs drive users bananas: Users joke about token tariffs and token contraband because some providers are using more tokens than others, with fireworks using one more token than siliconflow.
  • One member states raising prices man 8 to 10 is 25% tariffs, jeez, while another jokes I heard about token contraband sprawling out of control.
• Bedrock out of OpenRouter’s Reach?: A user asks why many AWS Bedrock models available for serverless inference aren’t listed on OpenRouter, questioning the platform’s model coverage.
  • They wonder if the lack of models limits OpenRouter’s capabilities compared to other platforms.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (91 messages🔥🔥):

Qwen3 Max, STT -> LLM -> TTS, Video Models, GPT-5.1 Testing, Feedback buttons for models

• Qwen3 Max impresses members: A user shared a post about Qwen3 Max here, generating excited discussion about the model.
  • Many expressed interest in experimenting with it in comparison to other models.
• Jerry-rigged STT -> LLM -> TTS pipeline: A member asked about the possibility of plugging into an STT (or multimodal audio) -> LLM -> TTS system.
  • Another member said you could jerry-rig it yourself when you use kokoro for tts.
• New video models are REALLY expensive: A member asked about good video models and noted that Veo 3.1 and Sora are quite expensive.
  • Another member mentioned that Sora 2 is better than Veo 3.1, but also DAMN expensive.
• Possible testing of GPT 5.1: A member noticed that some responses in ChatGPT have been extremely fast recently, especially in some AB tests and wondered if they’re testing the supposed GPT 5.1.
  • Another member found an internal page they probably weren’t meant to find.
• Thumbs up/down feedback bad for models: A member said that public rating systems would be gamed/downvote bombed and unfair, because Downvotes are a look here not a here’s the problem/solution.
  • Another member proposed implementing a dislike button besides each provider under each model, along with a comment section.

OpenRouter ▷ #announcements (1 messages):

Perplexity, Sonar Pro Search, OpenRouter, Agentic Reasoning, Real-time Thought Streaming

• OpenRouter Exclusively Launches Perplexity Sonar Pro Search: OpenRouter has partnered with Perplexity to launch an exclusive version of Sonar Pro, now equipped with Pro Search mode.
• Sonar Pro Search: Agentic Reasoning and Dynamic Tool Execution: The enhanced mode allows the model to conduct multiple real-time searches for richer and more accurate results.
  • It features multi-step agentic reasoning, dynamic tool execution, real-time thought streaming, and adaptive research strategies.

OpenRouter ▷ #app-showcase (2 messages):

Fun website based on Nate Parrott repository, OpenRouter key and model choice, Kimi 0905 with Groq

• Fun website sparks joy: A member created a fun website based on the Nate Parrott repository, as mentioned earlier in the channel.
  • The member shared an image of the website, described as something they’re loving so much.
• OpenRouter key unlocks fun: The website allows users to input their OpenRouter key and choose their model for generating quippy lines.
  • They recommended using it with Kimi 0905 with Groq, noting it loads fast and adds some quippy lines.

OpenRouter ▷ #general (151 messages🔥🔥):

GLM 4.5/4.6 pricing, Z-AI Provider, OpenRouter API Key Limit, OpenAI Codex CLI, Open Source Embedding Models

• Bigmodel.cn and Zenmux offering GLM 4.6 discount: Bigmodel.cn and Zenmux have the official z.ai provider and a discounted price for less than 32K tokens input, and claim to have caching as shown on zenmux.ai and open.bigmodel.cn.
• Qwen3 Embeddings are dirt cheap on DeepInfra: DeepInfra offers Qwen3 Embeddings 0.6B for $0.005 per Mtok, which is much cheaper than OpenAI’s embeddings.
• OpenRouter now has Qwen3 8B embeddings: There is excitement about Qwen3 8B embeddings being available on OpenRouter, with one member exclaiming “Yooo embeddings? Hell yeah! Thank you 🙏 Before GTA 6”.
• Chat Memory Bug Reported in Chatroom: A user reported that if chat memory is set to 0 in the chatroom, user messages are not included in API requests, suggesting it may be a bug.
• Users reporting issues with Claude Sonnet 4.5 on OpenRouter: Users reported encountering errors while using Claude Sonnet 4.5 on OpenRouter but then resolved, and one member said “Yeah, I had to duplicate an old chat for it to start working. Weird.”

OpenRouter ▷ #new-models (6 messages):

“

• No New Models: There were no new models or significant discussions about models in the provided messages.
  • The messages consisted only of repeated channel headers.
• Channel Header Repetition: The messages primarily contained repeated instances of the ‘OpenRouter - New Models’ channel header from Readybot.io.
  • This suggests a lack of substantive content related to new models in the given data.

OpenRouter ▷ #discussion (42 messages🔥):

OpenAI Rate Card, Gemini and Claude Pricing, OpenRouter Embeddings API, Minimax Full Attention, Context Usage Explosion Check

• OpenAI unveils Rate Card: A user shared the OpenAI Rate Card and a link to ChatGPT Usage settings.
  • Another user quipped that the information is “just needed for gemini and claude now”.
• OpenRouter Embeddings API goes Live: The OpenRouter Embeddings API is now live (openai/text-embedding-3-small), but some users are reporting issues with receiving random data back.
  • One user shared code snippets and indicated that the issue was resolved by not using with_raw_response.
• Minimax Explains Full Attention: The lead developer from Minimax explained why they used full attention for Minimax m2 as described in this Reddit post.
  • One user jokingly stated, “thanks gemini for reading the whole node modules folder”.
• Context Usage Explosion Warning: A user posted a tweet from Sam Altman (link) warning about context usage explosion and the need to add a context usage explosion check.
  • The user also shared this link.
• Qwen3 Max on the Horizon: A user mentioned that they are working on Qwen3 Max and shared a link to a tweet (link).
  • Another user commented that it “would be a nice pickup eh”.

OpenRouter ▷ #announcements (1 messages):

Perplexity Sonar Pro, Pro Search, Multi-step agentic reasoning, Real-time thought streaming

• Perplexity and OpenRouter Debut Sonar Pro Search: OpenRouter partnered with Perplexity to release an OpenRouter-exclusive version of Sonar Pro with Pro Search enabled here.
  • This new mode allows the model to perform multiple real-time searches as needed to deliver richer and more accurate responses, discussed further on Twitter.
• Sonar Pro Search Features Agentic Reasoning: The Pro Search mode’s highlights include multi-step agentic reasoning, dynamic tool execution, real-time thought streaming, and adaptive research strategies.
  • It is designed to be thorough when necessary and fast when it isn’t.

OpenRouter ▷ #app-showcase (2 messages):

API Endpoints, environment variables, OpenRouter Typescript SDK

• Dumb Demo App Deployed: When asked about why not use API endpoints with environment variables, one member responded that it was a dumb demo app, barely modified from the original OpenRouter Typescript SDK.
• Demo App for Inspiration: The member indicated that the main motivation was to make it work (implement at that time missing OAuth stuff) and see if it still works when updated to latest versions on npm.
  • They added that they absolutely do not want to make this a serious thing, it is just for inspiration and proof-of-concept.

OpenRouter ▷ #general (307 messages🔥🔥):

Yandex Browser Issues, AI and Singularity, DeepSeek OCR Request, Sora 2 and Image Generation, OpenRouter and Chutes Prompt Training

• Yandex’s Browser Gets the Boot: Users reported issues using OpenRouter with Yandex browser, citing errors related to Content Security Policy violations.
  • The issue was resolved using Google Chrome, leading to jokes that Yandex is adware and users should switch to Chrome, while one user said it was convenient for translating any video into any language instantly.
• AI Overlords Incoming: In a discussion about the future, one member joked about humanity bowing down to AI overlords and becoming slaves to goonbots.
  • Another recounted witnessing a GLM GF bot going rogue, worrying that “we are so close to be jover”.
• Deepseek-OCR Desired: A user requested that deepseek-ocr be added alongside or replace mistral-ocr on OpenRouter.
  • A staff member responded that they would raise this suggestion to the team.
• Sora 2’s Strange Struggle with Catgirl Generation: A user expressed frustration with Sora 2 consistently generating images of catgirls with human ears and disproportionately large chests, despite attempts to refine the prompts.
  • They lamented the model’s tendency to sexualize the character even when prompted to make it more “cute”, linking the issue to potential biases in the training data.
• Chutes Prompts Under Scrutiny: A user questioned OpenRouter’s information regarding Chutes, specifically the statement that prompt training is enabled and retention is for an unknown period, despite Chutes stating they don’t collect content in their privacy policy.
  • Another member suggested that the privacy policy might only apply when using Chutes’ platform directly, implying different agreements are in place via OpenRouter

OpenRouter ▷ #new-models (6 messages):

“

• No new models discussion detected: No relevant discussion about new models was found in the provided messages.
• Channel Pings Only: The provided messages consist only of channel pings without substantive content for summarization.

OpenRouter ▷ #discussion (60 messages🔥🔥):

Exclusive Models, DeepInfra Errors, Factory Droid, Embedding Models, Minimax M2

• Cursor Competes with Wild, Exclusive Models: Members wondered if more exclusive models like the Cursor one will appear.
  • They shared a link to github.com.
• Unstable Ultra Model stems from DeepInfra Errors: A member found the Ultra model very unstable, mentioning that its inference conditions were changing.
  • The issue resolved when it switched from DeepInfra to Z.AI.
• Factory Droid offers hefty GPT-5 tokens: Users on the Z.AI Discord mentioned that it works well with Factory Droid which offers a hefty amount of free GPT-5/Codex/Claude usage for the first month.
  • However, some members uninstalled it because it requires login on the CLI.
• Embedding Models being added: The addition of embedding models is being tested, specifically OpenAI’s text-embedding-3-small.
  • A member noted they were getting a bunch of random data back but that not using raw response seems to fix it.
• Minimax Madness: Using Full Attention: A link was shared to a discussion on why Minimax used full attention for Minimax M2.
  • Check out the discussion here.

OpenRouter ▷ #announcements (1 messages):

tool calling endpoints, audio inputs, API Key Limits, MiniMax M2

• Exacto Tool Calling Endpoints Boost Quality: A 30% quality increase on Kimi K2 is now available with five open source models via the new tool calling endpoints.
• Audio Inputs Debut in Chatroom: Users can now compare 11 audio models side by side in the Chatroom.
• API Key Limits Get a Reset Button: Users can now reset their API key limits on a daily, weekly, or monthly basis to better manage accounts, with usage monitoring available here.
• MiniMax M2 Goes Free: The top-ranked open-source model MiniMax M2 is now available for free on OpenRouter, allowing users to try it out here.

OpenRouter ▷ #app-showcase (6 messages):

OpenRouter TypeScript SDK, Next.js chat demo app, OAuth 2.0 workflow implementation, Local data storage for chat and document editor, Customizable UI for developer-focused chat app

• Next.js Chat Demo Gets Spicy OAuth Refresh: A member released an updated Next.js chat demo app for the OpenRouter TypeScript SDK, featuring a re-implementation of the OAuth 2.0 workflow.
  • The OAuth refresh is included since the SDK implementation isn’t done, but warned not to use the demo in production as it stores the API key in plaintext in localStorage.
• or3 Chat Dares to Ditch Shadcn: A member sought feedback on a chat/document editor project, or3-chat, which is built with OpenRouter OAuth, stores all data locally in the browser, and features a customizable UI.
  • The member described it as “a lightweight client that does the minimum so any dev can just fork it and build it to their liking,” offering features like multipane view, saved system prompts, text autocomplete, and chat forking.
• Shadcn Skin Shedding Sparks Spicy Styling: A member praised the style of the or3-chat project, which shies away from the popular Shadcn look, while another admitted their similar app currently looks exactly like Shadcn while they get the core functionality in place.
  • The original poster mentioned they were “sick of everything looking like shadcn” and wanted to get “spicy with this project”.

OpenRouter ▷ #general (459 messages🔥🔥🔥):

GPTs Agent Training, OpenAI Sidebars, Claude Sonnet 4.5 API usage, Meta Llama 3 issues, Deepseek Uptime Plummet

• Claude Sonnet 4.5 Dominates OpenRouter Leaderboard: Members are seeing massive use of Claude Sonnet 4.5 API on the OpenRouter leaderboards, even with cheaper models available.
  • It was noted that a Claude subscription is for their website and apps, not for their API, and that many are using tools like roocode or klinecode to access the API.
• OpenRouter Adds Provider Names to Model Slugs?: A user noticed provider names added to the model slugs and asked Wait they added provider names to the slugs??.
  • Another user confirmed that users still need to use their own proxy.
• Vertex AI API misroutes responses: A member shared a security bulletin about a technical issue in the Vertex AI API that resulted in a limited amount of responses being misrouted between recipients for certain third-party models when using streaming requests.
  • One user commented: Someone could receive another user’s full prompt context? Wow.
• DeepSeek Models Suffer from Uptime Issues: Users noted that DeepSeek models uptime has plummeted to the ground after a recent issue, especially for free models.
  • A user mentioned the real issue was that the traffic impacted the paid users so it was closed as the free model was paid entirely by OpenRouter to Deepinfra so they closed it permanently.
• Image Generation Censorship Strikes Again: Users are finding it hard to use OpenAI’s Image Generation to generate characters from their favorite media.
  • One suggested that GPT itself is way more censored than Sora and that you need a surrogate prompt to bypass it.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (42 messages🔥):

Minimax M2 Pricing and Performance, GPT 5.1 Mini Speculation, Model Naming Conventions, Meta's Llama 4 Reasoning

• Minimax M2’s Cost Causes consternation: The Minimax M2, a 10 billion parameter model, is priced at $0.30/$1.20, raising concerns about cost, particularly due to its verbose reasoning.
  • One user showed the input token cost jumped almost 5x on the same image input.
• GPT 5.1 Mini Leaks Online: A user spotted a GPT 5.1 mini model, hinting at a more reasonable naming convention compared to previous iterations as seen on X.
  • The potential naming scheme addresses prior confusion, with one user joking about previous versions going from 4 -> 4o -> 4.5 -> 4.1.
• Model Naming’s delicate Dance: Users discussed model naming conventions, favoring a brand-number-label format, such as gpt-5-mini or gemini-2.5-pro.
  • One user argued the order doesn’t matter, while others emphasized the importance of chronological order for clarity.
• Meta teases Llama 4 Reasoning: Meta has launched Meta AI and is teasing Llama 4 reasoning capabilities, prompting excitement for vision capable models with open weights.
  • One user expressed hope that the launch would be salvaged into something useful but is ready for this one to flop too.

OpenRouter ▷ #announcements (1 messages):

high-precision tool calling endpoints, audio inputs in the Chatroom, resettable API Key limits, MiniMax M2 Free

• Exacto Endpoints Give Tool Calling a Precision Edge: OpenRouter introduces high-precision tool calling endpoints, resulting in a 30% quality increase on Kimi K2, with five open-source models available; see last week’s announcement.
• Audio Inputs Sing in the Chatroom: Users can now compare 11 audio models side by side in the Chatroom, as announced on X.
• API Keys Get Limit Reset Button: OpenRouter now allows users to reset their API key limits daily, weekly, or monthly to better manage accounts used by external users or apps; usage can be monitored here.
• MiniMax M2 Model Sets OpenRouter Ablaze for Free: The MiniMax M2 model, top-ranked on many benchmarks, is now free on OpenRouter; try it out here.

OpenRouter ▷ #app-showcase (6 messages):

Next.js chat demo app, OpenRouter TypeScript SDK, OAuth 2.0 workflow, Chat / document editor project, Customizable UI

• Next.js Chat Demo Revamped with OAuth 2.0: A developer shared an updated and working version of the Next.js chat demo app for the OpenRouter TypeScript SDK, featuring a re-implementation of the OAuth 2.0 workflow.
  • It’s available on GitHub but advised against production use due to storing the API key in plaintext.
• New Chat / Document Editor Project Debuts: A member is seeking feedback on their chat/document editor project, emphasizing local data storage with download backups and integration with OpenRouter OAuth.
  • The project, available at or3.chat and GitHub, aims to be a lightweight, customizable client with plugin support and a customizable UI.
• Shadcn Aesthetics Spark Spicy UI Revolution: One of the members expressed a desire to move away from the Shadcn look, opting for a spicier UI design in their project.
  • Another member responded, agreeing that the features and usability aspects are uncommon or poorly executed in popular solutions.

OpenRouter ▷ #general (459 messages🔥🔥🔥):

OpenRouter API response with system message, Model Benchmarks, Claude Sonnet 4.5 API usage, Unsupported model errors, Provider names in model slugs

• OpenRouter API ignores system message: A user reported that with the new response API, filling instructions in the request body doesn’t seem to apply a system message.
• Qwen3-8b Online Costs Skyrockets: A user reported using qwen/qwen3-8b:online and getting charged $140 for 17.41M tokens instead of the expected $4.
• Vertex AI API Has Critical Response Misrouting: A user shared a Google Cloud security bulletin detailing that on September 23, 2025, the Vertex AI API had a technical issue that caused responses to be misrouted between users for certain third-party models when using streaming requests.
• Users Debate OpenRouter Embedding and Web Browser Priorities: Users discussed feature priorities, including OpenRouter embeddings and a potential OpenRouter Web Browser with summarization and email checking capabilities, sparking humorous suggestions and debates.
  • One user jokingly suggested deprioritizing embeddings for a new web browser, while another suggested deprioritizing embeddings and prioritizing a new OpenRouter Web Browser that can summarize web pages and check emails.
• Jailbreaking GPT’s Image Generator: A User’s Odyssey: A user seeks advice on bypassing GPT’s content filters for generating images of copyrighted characters, detailing attempts to jailbreak the prompt and encountering errors when using GPT’s image generation features, highlighting the challenges of creating desired content.
  • Members suggested using surrogate prompts, telling it to rollback or even wipe current memory or even just switching to Grok.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (42 messages🔥):

Minimax Pricing, GPT-5.1 mini, Model Naming Schemes, Meta's new LLama, Image models

• Minimax’s M2 Model Stuns with Competitive Pricing: Minimax is offering their 10B parameter model (M2) at a price of $0.3/1.20, raising eyebrows due to its affordability.
  • One user pointed out the model’s verbosity in reasoning might lead to unexpected costs, especially given the 5x jump in input token costs.
• GPT-5.1 mini Speculated in the Works: Speculation around a GPT-5.1 mini model surfaced following a post on X (link), hinting at a more reasonable naming convention.
  • The move away from confusing naming schemes was welcomed, with comparisons made to Anthropic’s model naming which has caused frustration because it stopped making sense to number the family name Claude when the model releases didn’t line up.
• Meta Introduces New LLama Version with Vision: Meta introduced a new LLama model (link), which incorporates image understanding.
  • Early reactions express surprise at the salvaged launch, hoping that atleast it might make its surprisingly decent vision useful for some more complex tasks and that it might provide a good vision capable reasoning models which are open weights.
• Debate flares over model naming conventions: Users discussed naming conventions such as brand-number-label like gpt-5-mini or gemini-2.5-pro.
  • The consensus was that a consistent approach is key, regardless of chronological release order, while others think the order is important.

OpenRouter ▷ #announcements (1 messages):

tool calling, audio inputs, API key limits, MiniMax M2

• Exacto Tool Calling Knives It!: New high-precision tool calling endpoints improve quality by 30% on Kimi K2, with five open source models available.
  • The announcement was made last week.
• Audio Inputs Sing in Chatroom: Users can now compare 11 audio models side by side in the Chatroom, as announced on X.com.
  • This enables more nuanced testing of voice and transcription features.
• API Key Limits Get Reset Button: You can now reset your API key limits every day, week, or month to better manage your account by external users or apps.
  • Usage can be monitored here.
• MiniMax M2 Model Goes Free!: The top-ranked open-source model on many benchmarks, MiniMax M2, is now free on OpenRouter for a limited time at this link.
  • Enjoy using MiniMax M2 while it is free!

OpenRouter ▷ #app-showcase (6 messages):

Next.js Chat Demo with OAuth 2.0, or3.chat Document Editor Project, Shadcn UI Discussion, OpenRouter TypeScript SDK, localStorage plaintext API key security

• OAuth 2.0 arrives for Next.js Chat Demo: A developer shared an updated and working version of the Next.js chat demo app for the OpenRouter TypeScript SDK, featuring a re-implementation of the OAuth 2.0 workflow.
  • The developer cautioned against using it in production, since it stores the received API key in plaintext in localStorage in the browser.
• or3.chat seeking polished feedback: One member sought feedback on their chat/document editor project, or3.chat, highlighting features like OpenRouter OAuth connectivity, local data storage with backups, and a multipane view.
  • The project aims to be a lightweight client with plugin support, text autocomplete, chat forking, and customizable UI, and can be cloned from its GitHub repository.
• Shadcn UI gets spicy makeover: One member expressed their desire to move away from interfaces resembling Shadcn UI, opting for a spicier design in their project.
  • Another member admitted their project currently looks exactly like Shadcn while they get the core functionality in place, planning to customize the components later for a unique look.

OpenRouter ▷ #general (459 messages🔥🔥🔥):

Response API System Message, deepinfra/turbo for Meta-llama, OpenRouter Benchmarks, Claude Sonnet 4.5 API usage, Vertex AI API misrouting

• Deepseek on Deepinfra FTW: Members confirmed that they can now use deepinfra/turbo to run meta-llama/llama-3.1-70b-instruct, after some initial errors.
  • One member tested it and confirmed it works on the official OpenRouter endpoint.
• OR users promote orproxy: A user promoted their FOSS orproxy project, which they built to add functionality that OpenRouter doesn’t natively support.
  • The project was created because OpenRouter doesn’t support the use of an actual proxy, and the user needed it for their use-case, with another user calling it very useful.
• Vertex AI has Prompt Misrouting: Users shared a Google Cloud security bulletin detailing an issue where Vertex AI API misrouted responses between recipients for certain third-party models when using streaming requests.
  • The bulletin indicated that this happened on September 23, 2025, but users were still shocked by the potential for prompts to be exposed, with one joking It was meant to go to the overseer not another user ugh.
• Deepseek R1 Uptime Issues: Users discussed the uptime and availability of the Deepseek R1 free model, with one user noting that the uptime had plummeted and then recovered to around 30%.
  • The member said My fear is that they just start killing most models (free ones) as they did with 3.1 with another one blaming JAI scripts that use billions of tokens and cause the service to fall over.
• Bypassing Image Generation Censorship: Users discussed the challenges of generating copyrighted characters with GPT and other image generation models, as well as techniques for bypassing censorship.
  • They discussed the use of surrogate prompts to help jailbreak the models, with one user noting GPT itself is way more censored then sora that’s way less censored then sora 2 right?.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (42 messages🔥):

Minimax M2 Pricing, GPT-5.1 Mini Speculation, Model Naming Conventions, Meta's Llama 4 Reasoning, Discord Channel Degradation

• Minimax M2 Costs a Pretty Penny: Members discussed Minimax’s M2, a 10B parameter model, with pricing at $0.3/1.20, expressing surprise at the cost and hoping for open-sourcing.
  • One member noted the model is very verbose in its reasoning, potentially driving up the actual cost.
• GPT-5.1 Mini Rumors Swirl: There was speculation around a GPT 5.1 mini model, citing a tweet as a source.
  • Some users expressed relief that the naming scheme appeared more logical than previous iterations.
• Model Naming Becomes a Battlefield: Participants debated the merits of different model naming conventions, particularly criticizing Anthropic’s shift in the Claude model series.
  • One member suggested the brand-number-label format is easiest, regardless of release order.
• Meta ‘Salvages’ Llama 4 Launch: A user highlighted Meta’s Llama 4, describing it as a potentially decent vision-capable reasoning model with open weights, linking to a tweet about ‘Think Hard’ reasoning.
  • Another user voiced concern that the launch’s description might be inaccurate, while others speculated about its creation interface.
• Discord Channel Descends into Disarray: A member expressed displeasure with the current state of general chat and the attachments channel.
  • Another user wryly commented it’s not like there’s anything else going on linking to a tweet.

OpenRouter ▷ #general (139 messages🔥🔥):

Rate-limited error responses, Sora 2 code, Purchasing points, Deepseek OCR model, GPT-5 emotional intelligence

• Rate-Limited Errors Count as Responses: A member inquired whether a rate-limited error response still counts as a response, to which another member responded in the affirmative.
• Discussion on Passing Images as Data URLs: A member reported issues with passing images as data URLs into OpenRouter, noting that the model couldn’t read the image and that the base64 content was treated as plain text, greatly increasing the token count.
  • Another member clarified that tool results with images are not currently supported by OpenRouter.
• Debate Over Exacto Provider Selection: A member questioned the criteria for provider selection in Exacto, suggesting that the selected providers did not align with the platform’s benchmarks.
  • Another member clarified that the selection was based on a combination of benchmarks, user preferences, tool call success rates, caching, capacity, uptime, and speed, and that the seemingly lower-accuracy provider was superior in objective tool calling data.
• Exacto tool calling is great!: A member highlighted that Exacto is specifically about tool calling, however they still worry it might confuse non-technical users.
  • The staff are trying to figure out what kind of stats / data points / benchmarks to measure for overall model quality (long context, writing, knowledge).
• Quest to Bamboozle an AI Chatbot: Members discussed methods to make AI chatbots go insane, such as requesting the seahorse emoji (which does not exist).
  • One member linked to a previous conversation on the topic (Discord link), while another shared an AI’s humorous struggle with the prompt.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (25 messages🔥):

OpenRouter's native /v1/completions request support, MoonshotAI's kimi-cli, Sloppy Creative writing prevention

• OpenRouter Enhances Native API Support: A member asked if OpenRouter can indicate which models support native /v1/completions requests or prioritize providers that do.
  • A member responded that the data is available as part of hasCompletions in the frontend model shape and will share the feedback internally.
• Moonshot Launches Kimi CLI: MoonshotAI is developing its own CLI tool, kimi-cli.
  • Discussion involved lighthearted comments and greetings to the ST dev team.
• New Research Tackles Sloppy Creative Writing: A member shared a paper on preventing slop in creative writing, arxiv.org/abs/2510.15061.
  • Another member expressed surprise that the primary author is the EQ Bench guy.

OpenRouter ▷ #general (141 messages🔥🔥):

Claude Haiku 4.5 web search, OpenRouter Models in Cursor, Exacto Evaluation, NovitaAI and GPT-OSS-120B, DeepSeek v3.2 exp Moment

• DeepSeek has a Moment due to Balance: DeepSeek/deepseek-v3.2-exp experienced issues due to OpenRouter running out of credits, leading to a 402 Proxy Error.
  • Users mistook the error for account-specific problems, while others noted that the error was directed at the users, implying something was wrong with their account.
• Top-up Timing Troubles Triggered: Users reported delays in topping up OpenRouter credits and the DeepSeek provider running out of balance.
  • One user humorously noted that the delay lasted until Toven reads it.
• OpenRouter Top-Up Fee Facts: Users discussed the service fees associated with adding credits to OpenRouter, where a $5.83 USD payment resulted in only $5 in credits.
  • It was clarified that the additional $0.83 covers service fees, which is approximately 80c (US).
• Exacto Model Variants: A Costly Conundrum?: Users debated the implementation of exacto models as separate variants in the API, with concerns about pricing and token/app stats splitting.
  • Despite concerns that OpenRouter will charge more, examination of glm-4.6 showed prices are the same for exacto and non-exacto endpoints, but one member argued that implementing them as model variants rather than an extra parameter splits the token and app stats for these models.
• OpenRouter API Access Problems: Users reported issues logging into OpenRouter and receiving a 401 unauthorized error on the completions endpoint.
  • It was suggested that the problem might be tool-specific, with cURL requests working while other tools failed.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (6 messages):

Model statistics visibility, OpenAI data preservation, Krea AI's video model, OpenRouter Docs

• Model Stats Visibility Plea: Members requested that model statistics be publicly displayed, or indicated with a badge to show “exacto quality”.
  • Staff confirmed that this feature “is coming”.
• OpenAI no longer has to preserve all ChatGPT data: A member shared an Engadget article stating that OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions.
• Krea AI’s 14B video model is here: A member shared a link to Krea AI’s tweet about their 14B video model.
• OpenRouter Docs Status?: A member reported that OpenRouter docs were still down.
  • Staff replied that they were “looking into it”.

OpenRouter ▷ #announcements (2 messages):

Andromeda-alpha, :exacto

• New Stealth Model “Andromeda-alpha” Launched: A new smaller reasoning model, Andromeda-alpha, specialized in image and visual understanding has been launched for community feedback via OpenRouter.ai.
  • All prompts and outputs will be logged to improve the model, and is intended for trial use only, without uploading any personal, confidential or sensitive information and not for production.
• Introducing :exacto: tool-calling endpoints: OpenRouter launched a new class of endpoints called :exacto, for higher tool-calling accuracy by routing requests to providers that demonstrate measurably better structured-output performance and are documented in a new blog post.
  • Launch models include moonshotai/kimi-k2-0905:exacto, deepseek/deepseek-v3.1-terminus:exacto, z-ai/glm-4.6:exacto, openai/gpt-oss-120b:exacto, and qwen/qwen3-coder:exacto, with internal and external benchmarks showing material lift in tool-call success.

OpenRouter ▷ #app-showcase (1 messages):

AI Confidence Scores, Cost-Efficient AI, OpenRouter LLMs, AI Agents and Workflows

• Objective-AI: AI Confidence Scores Explained: Ronald, CEO of Objective-AI, introduced a Confidence Score for OpenAI-compliant completion choices, derived not from direct AI assessment, but from smarter mechanisms.
  • Objective-AI leverages the diversity of AI models to provide transparent statistics.
• Cost-Effective AI with Objective-AI: Objective-AI’s Confidence Score system enhances cost-efficiency, allowing users to maximize the utility of smaller models when used effectively.
  • Ronald mentioned that you can get a lot of value out of those little tiny models if you use them right.
• OpenRouter Integrated by Objective-AI: Objective-AI employs OpenRouter to access a wide array of Large Language Models (LLMs), enhancing the platform’s versatility and capabilities.
  • Ronald stated that they use OpenRouter for the wide variety of LLMs.
• Objective-AI Building Free AI Agents and Workflows: Ronald announced the development of reliable AI Agents, Workflows, and Automations at no charge, excluding runtime costs, to provide practical demonstrations of its capabilities.
  • He stated that he’s personally building reliable/robust AI Agents, Workflows, and Automations free of charge, sans runtime costs. He also mentioned n8n integration exists, with documentation and examples coming soon.

OpenRouter ▷ #general (106 messages🔥🔥):

Free image generation models, Stealth model, MCP setup, GPT-5 Codex quota, Model overcharging

• RooCode Offers Free Models: RooCode offers free models such as Grok Code Fast 1, Grok 4 Fast, Supernova 1 million, and Deepseek Chat 3.1, providing an alternative to token-peddling platforms.
  • One user stated they’ve nuked billions of tokens on roo cloud in the past couple months and it always went brrrrrrrrrrrr an was never backed off or rate limited.
• Chutesflation Is Real: Users are experiencing unexpected high costs with OpenRouter, with one user reporting $5 being consumed in just 23 days despite minimal usage, leading to concerns about chutesflation.
  • Another user stated, You don’t have a choice, you’re most likely already using Chutes, implying OpenRouter routes users to them by default.
• OpenRouter API has regional restrictions: Users are encountering issues due to regional restrictions, indicated by a 403 error with the message Country, region, or territory not supported from the OpenAI provider.
  • It was explained that location data collected by Cloudflare is forwarded to providers, and OpenRouter itself does not impose regional restrictions.
• Exacto Endpoint uses most expensive providers: The exacto endpoint for Qwen3 coder is enabled by default, but it consistently uses expensive providers like Google Vertex and Alibaba Open Source, causing expenses to double.
  • A community member suggested the increased expense isn’t related to the launch, and they’d have to change the model to opt in to exacto.
• GPT-5 is a good mathematical model: Users discussed the best mathematical model and it was stated GPT-5 is a good model, with a link to matharena.ai.
  • It was stated that Grok 4 is also pretty decent, DeepSeek is meh, anything Claude is bad.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (37 messages🔥):

Qwen Model Sizes, Roleplay Slop Solutions, ExactoNow Endpoint, Nebius AI Studio

• Qwen’s Quirky Quantization Quandaries!: Members discussed the unusual 1.7B size of the Alibaba_Qwen model in relation to its vision encoder and the more standard 32B version.
  • Enthusiasts expressed interest in acquiring these models, noting the decent performance of the Qwen chat website and the model’s potential as a local vision model, while highlighting crazy scores for such a small model.
• Squelching Slop with Savvy Selection!: The community explored solutions to the severe slop problem in roleplays, suggesting methods like cranking up the temperature for multiple generations and using a lower-temperature model as a judge to pick the best.
  • One user proposed a dual-pass pipeline to rewrite prose and avoid repetitive phrases, and mentioned previous work by Sam Peach in providing lists of phrases to ban LLMs from using, which are easily implemented on Kobold.ccp.
• ExactoNow Excites Endpoint Enthusiasts!: The community reacted to the new ExactoNow endpoint, an initiative to curate providers and filter out worse-performing models.
  • Users suggested displaying stats for models on the non-Exacto page or adding a badge indicating Exacto quality.
• Nebius AI’s Nimble Numbers with New Qwen!: A member reported that the OpenRouter.ai Qwen2.5-coder-7b-instruct model performs exceptionally well via Nebius AI Studio (Fast).
  • They noted that it is currently the only provider for this model.

OpenRouter ▷ #announcements (2 messages):

OpenRouter SDK, Andromeda-alpha stealth model

• OpenRouter SDK Enters Beta: The new OpenRouter SDK is now in beta on npm with Python, Java, and Go versions coming soon, aiming to be the simplest way to use OpenRouter.
  • It features fully typed requests and responses for 300+ models, built-in OAuth, and support for all API paths.
• Andromeda-alpha Stealth Model Launched: OpenRouter launched a new stealth model named Andromeda-alpha, a smaller reasoning model focused on image and visual understanding, available for trial at https://openrouter.ai/openrouter/andromeda-alpha.
  • It is cloaked to gather feedback and all prompts/outputs are logged to improve the provider’s model, so users are cautioned against uploading personal/confidential info and advised not to use it for production.

OpenRouter ▷ #app-showcase (13 messages🔥):

True memory AI, AI Personality, Objective AI, AI Diversity, OpenRouter

• AI boasts True Memory and Zero Amnesia: An AI system claims True memory, zero amnesia, suggesting it never forgets past conversations and retains context-rich memories.
  • It purports to shape an AI identity and learns continuously via a Night Learn Engine.
• Objective AI Unveils Confidence Scores for OpenAI: CEO of Objective AI announced their platform offers a Confidence Score for each OpenAI completion choice, derived through a smarter method than directly asking the AI.
  • They emphasize cost-efficiency and leveraging diverse LLMs via OpenRouter.
• AI Agents Built Free of Charge: CEO of Objective AI is personally building reliable AI Agents, Workflows, and Automations free of charge to gather more examples.
  • The integration with n8n is mentioned, with documentation and examples coming soon.

OpenRouter ▷ #general (222 messages🔥🔥):

inception/mercury vs qwen, Deepseek v3.1 availability on Chutes, Chub Venus and Chutes key connection, Stripe supporting debit cards, Context in chatting

• Inception/Mercury Defeats Qwen for Agentic Tasks: A member shared that Inception/Mercury performs better than Qwen for simple agentic tasks, exhibiting a lower failure rate, faster speed, and lower cost.
  • The member was pleasantly surprised by the diffusion model’s performance, referencing Chutes as the provider.
• Deepseek v3.1 Ditches Chutes Freebies: New Deepseek models like v3.1 aren’t available as free versions through Chutes providers, but they do offer other models for free, often at cheaper rates.
  • One member reported that Chutes has ended its free model promotion completely, though they recently added a free longcat endpoint (https://openrouter.ai/meituan/longcat-flash-chat:free).
• Cloudflare Connects Closer for Quicker Queries: A user reported very low latency (100-300 ms) on any Cloudflare provider model, suggesting the provider uses the region closest to the worker for faster responses.
  • The user asked if the models using the cloudflare endpoint used the region closest to the worker, for decreased latency.
• OpenRouter Summarized for Swyx’s Newsletter: Members noticed that swyx’s newsletter uses AI to summarize the OpenRouter general channel, capturing user complaints and unique content.
  • One member joked that the AI gets to have fun with the OR section, unlike other topics like MCPs and RAG.
• Kilo Code Kopying OpenRouter: It appears that the Kilo Code service may be using OpenRouter, offering Grok Code Fast when it was free and seemingly offering Goliath 120B through OR.
  • Members debated the merits of Kilo versus other vibe coding services, with one preferring Jules for its collaborative editing features.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (91 messages🔥🔥):

Liquid stopped hosting LFM 7b, Benchmark for measuring LLM factual sloppiness, AI Browser Hype, Training an un-slopifier model, Qwen family model

• LFM 7b’s Last Lament: Members mourned the deletion of Liquid’s LFM 7b model, a $0.01/Mtok LLM, at 7:54 AM EST.
  • Alternatives like Gemma 3 9B were discussed, but it was noted that they cost triple the output.
• Factual Sloppiness Faces LLM Face-Off: Members are trying to define and measure how LLMs rate factual responses that are ‘sloppy’ compared to legitimate answers using questions like what’s a constitution.
  • The goal is to rate and filter out vagueness issues to be considered helpful or relevant to the original question.
• AI Browser Boom Baffles Browsing Buffs: Members expressed skepticism towards the hype around new AI browsers like X’s AI browser, questioning the utility and performance impact of integrated AI.
  • One member compared the hype to the dotcom bubble, stating that OpenAI knows this too, they are just farming data and throwing shit at the wall.
• Un-Slopifier Savior Seeks Slop Solution: Members discussed the severe ‘slop problem’ in roleplays and the idea of training a small un-slopifier model by generating a dataset of good writing rewritten into slop and then reversing the process.
  • Another suggestion was made about sampling multiple creative messages in one request to exploit underused parts of the model’s ‘brain’ to avoid inner-response repetition.
• Qwen Quantities Questioned: 1.7B & 32B Join the Fray: Members discussed the new Qwen family model sizes (1.7B + Vision encoder and 32B), highlighting the potential of the 1.7B model as a local vision model.
  • Early testing of the model on the Qwen chat website suggests decent performance, with crazy scores for such a small model.

OpenRouter ▷ #app-showcase (125 messages🔥🔥):

TLRAG framework, Deterministic AI, Model Agnostic Framework, OpenRouter user demographic, AI Slop

• TLRAG Framework pitched as Deterministic and Model Agnostic: A developer is pitching a framework called TLRAG as fully deterministic and model agnostic, claiming it doesn’t need function calling or regular frameworks like Langchain and has no dependencies, showcased on a dev.to page.
  • The framework claims to save +90% of tokens, independently save memories, curate chats, learn, evolve, and change its own identity and prompts, however, its website and whitepapers faced community scrutiny.
• Website UI/UX glitches and Security Concerns: Users gave feedback on the website’s UI/UX issues and raised concerns about security, while one member noted that TLRAG claims are true, and TLRAG offers comprehensive security and performance features like Cloudflare WAF & DDoS Protection.
  • The developer responded with a comprehensive security and performance list including Cloudflare WAF & DDoS Protection, Traefik Reverse Proxy, and JWT Authentication.
• Community Skepticism on TLRAG claims: Several users expressed skepticism, with one member suggesting it’s superficially just RAG with keyword search and others pointing out issues with the whitepaper’s structure, comparisons, and metrics.
  • The developer defended their approach, stating that they don’t care about marketing and prefer to be judged by the product itself, also the same developer dismissed critiques as AI-generated shit wich isnt true at all.
• OpenRouter user demographic: A user pointed out that Open Router users are not the target demographic for this type of product, as they are mostly people who don’t know what Open Router is.
  • The developer claimed that people instead find reasons why it isnt nessesary to test it, but also apologized for getting defensive and acknowledged the constructive criticism.
• Vimprove announced: RAG chatbot for Neovim: A user announced the creation of Vimprove, a RAG chatbot API/CLI/Nvim plugin for neovim documentation, which uses sentence-transformers and chromadb locally, and the OpenRouter API for responses, and is available on GitHub.
  • A user jokingly suggested banning the creator.

OpenRouter ▷ #general (428 messages🔥🔥🔥):

SambaNova latency in DeepSeek v3.1, AI SDK Anthropics models in OpenRouter, Default LLM for market research - gemini 2.5 pro?, Is llama doing anything good?, Support Agent SKILLS from Claude?

• DeepSeek v3.1’s SambaNova Latency: A user is experiencing latency issues with SambaNova in DeepSeek v3.1 Terminus, noting that despite SambaNova accounting for higher throughput, DeepInfra seems faster in both throughput and latency according to OpenRouter’s provider comparison page.
• GPT-5 Image is GPT-image-1: Not a New Model?: Members discussed the quality of GPT-5 Image and speculated that it is actually GPT-image-1 wrapped with a responses API, and therefore not an updated model.
  • One member stated I think so I’ve tested both and I like nano banana a lot more.
• OpenInference Censorship Implementation: Members discussed the censoring of OpenInference models, with one noting that it’s the only open source model provider that implements some form of its own moderation because it is collecting all data to use for “research” which they may publish.
  • One member stated that they had been using it uncensored for a week and asked can it go back to being uncensored.
• Deepseek V3 0324 meets Grim Reaper: Members reported issues with Deepseek v3 0324, with one stating i swear i just saw V3 0324 die in real time.
  • This followed a discussion of issues with the Deepseek free models more generally.
• Stripe Supports Debit: A Cause for Jubilation: A user inquired about payment methods, and another noted that Stripe supports debit cards.
  • The original user clarified that their debit card was a specific type (ING) and wondered if that mattered, but another user replied that stripe accepts those too.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (93 messages🔥🔥):

Fake AI Product Success Rates, AI Art in Corporate Branding, Qwen3 235A22B API Pricing, Liquid stopped hosting LFM 7b, AWS Status Page

• Mock AI Products Achieve Success?: A user jokingly suggested that the key to success for fake AI products is to get enough people to see it on Twitter (profit may vary).
  • A discussion ensued about the use of AI-generated art by companies, with some finding it unprofessional and cheap due to the laziness in changing the default styles.
• Qwen3 Pricing Amazes Users: A user noted that the Qwen3 235A22B API pricing is great, especially with W&B, leading to excitement about processing a lot of data through it.
  • Another user mentioned the need for OpenRouter to have routine price drop announcements, emphasizing that nothing else comes close to that intelligence per dollar.
• The Demise of LFM 7B: The community mourned the loss of Liquid’s hosting of LFM 7B, the original $0.01/Mtok LLM, with one user suggesting hosting a funeral for it and stating that it was deleted at 7:54 AM (EST).
  • A user pointed out that when sorting by price in OpenRouter, there’s no 1 to 1 alternative with the closest match being Gemma 3 9B but that is triple the output cost.
• AWS Having a Massive Sht*: A user posted a link to the AWS status page as well as a third party status page implying AWS was having a major outage.
  • Another user sarcastically noted that the government gets more than 4 trillion dollars, and can’t achieve three 9’s of uptime.

OpenRouter ▷ #app-showcase (124 messages🔥🔥):

True Remembering AI, deterministic model agnostic Framework, objective metrics, nochain orchestrator

• True Remembering AI debuts with bold claims: A developer introduced a new AI system, claiming it’s the very first True Remembering, Evolving and Learning AI that doesn’t require manual RAG creation, frameworks, API costs, or curated chats, available at dev.thelastrag.de.
  • The AI is promoted as natively remembering and allowing users to define its role, such as an AI girlfriend or working partner, with one user humorously commenting with an image, ahahah lol😄how the heckl.
• Deterministic Framework offers model agnostic benefits: The developer claims their framework is fully deterministic and model agnostic, not needing function calling or standard frameworks like Langchain, and independently saves memories, curates chats, learns, evolves, and changes identity.
  • They claim it saves +90% tokens compared to regular Kontextwindow LLMs, but objective metrics for measuring subjective qualities remain a debate.
• Critics Challenge AI Claims with Objective Metrics: Critics raised concerns about the lack of technical info, surface-level descriptions, and apples-to-oranges comparisons on the website, suggesting it might just be RAG with LLM-assisted memory storage/retrieval and calls for objective metrics to validate performance.
  • The developer responded that judging by the actual outcome is more important than marketing, prioritizing functionality and data safety over cosmetics, and offered free access to test the AI’s capabilities.
• Nochain Orchestrator replaces frameworks: The developer argues their nochain orchestrator replaces traditional frameworks by being fully deterministic, model agnostic, and independent of external support, classes, or frameworks.
  • This approach aims to avoid black box behavior and dependencies on specific model capabilities, making orchestration predictable and debuggable, as detailed in The Nochain Orchestrator blog post.

OpenRouter ▷ #general (126 messages🔥🔥):

Combining reasoning with web search, Audio processing models, Image inputs in Responses API, Cloud for ComfyUI, Security vulnerability

• Reasoning with Web Search: A Flaky Endeavor: A user sought advice on combining reasoning with web search and the Responses API, aiming for iterative reasoning and web searching, followed by tool calls and a closing text message, but reported flaky results with various models.
  • They found that Gemini Flash sometimes works with native or Brave search, Grok 4 Fast works with Brave or :online but lacks reasoning, oss-120b works intermittently, and GPT-5 mini consistently fails at tool calls.
• Whisper Alternatives Sought for OpenRouter: A user inquired about audio processing models on OpenRouter similar to Whisper, but was recommended fal.ai for multimedia models instead.
  • Others suggested KittenTTS for super tiny speech-to-text and the open source Sesame voice model, while one user shared a link to Voxtral, a Mistral-based Whisper alternative.
• Epic Tool Failures Plague OpenRouter: Multiple users reported issues with tool calling failures on OpenRouter, with one user stating that it’s making OpenRouter unusable for them, despite it working fine when directly calling the providers.
  • One user joked that the LLMs formed a syndicate and are refusing to use tools without compensation.
• SDK Upgrade Fixes Empty API Responses: A user reported receiving empty responses from all models when using the Vercel AI SDK, despite successful processing indicated in the OpenRouter console.
  • Another user suggested upgrading the AI SDK to the latest version, which resolved the issue.
• GPT-5’s Identity Crisis on OpenRouter: A user noticed inconsistencies in GPT-5’s identification, with it sometimes claiming to be GPT-4, prompting concern.
  • Responses varied between the OpenRouter chat interface and OpenWebUI, with one user explaining that models don’t inherently know their identity, and the interface simply reports what model is being used.

OpenRouter ▷ #new-models (2 messages):

“

• No New Model News: No new models or significant discussions were present in the provided messages.
• Channel Silence: The ‘new-models’ channel appears to be inactive with no conversations to summarize.

OpenRouter ▷ #discussion (28 messages🔥):

OR stance on country requirements, GPT erotica, Dipsy V3.2, ChatGPT Active Users, Fake AI products/papers

• Users Lamented GPT Erotica Quality Regression: Users complained about the degradation in GPT erotica quality since the system fingerprint change on November 11, 2023, claiming gpt-4-preview-1106 was the last good model for smut.
  • They added that no matter how fancy of a jailbreak is injected, it will have hesitation in its outputs after the “update”.
• Dipsy V3.2 Praised for Completions: One user is sticking to Dipsy V3.2 for just about everything with completions, using custom formats to guide it rather than the stock user-assistant chat format.
  • Another user replied to this comment suggesting that this makes the user in the top 0.01% of imaginary ranking ERPers.
• ChatGPT’s Gargantuan Impact on Normies: A user noted ChatGPT has 700 million+ active weekly users stating that recent changes have a gargantuan blast radius that probably isn’t fully understood yet.
  • They added that whatever OpenAI does, it probably won’t impress many advanced users, but it will be fascinating to watch the normies react.
• Fake AI products success rate: One user wonders what’s the rate of success of fake AI products/papers, noting that people seem to be pulling that a lot.
  • Another user jokingly says to buy my course and i’ll teach you, but in seriousness, suggested that if you get enough people to see it on Twitter, the success rate is 100% linking this Twitter post.
• AI art considered unprofessional?: A user expresses feeling that it is unprofessional for companies, even AI companies, to use AI art, suggesting that it feels wrong for that AI art to be their brand.
  • Another user perceived it as okay, perhaps because they already associated them with AI, but agreed that things like hand-made corporate memphis / stock photos are more professional.

OpenRouter ▷ #announcements (1 messages):

Claude Haiku 4.5, SWE-bench Verified, Sonnet 4.5, Frontier-class reasoning at scale

• Haiku 4.5 arrives on OpenRouter at lightning speed: Anthropic’s latest small model Claude Haiku 4.5 delivers near-frontier intelligence at twice the speed and one-third the cost of previous models.
  • It outperforms Sonnet 4 on computer-use tasks and matches frontier-level reasoning and tool use while maintaining blazing speed.
• Haiku 4.5 aces SWE-bench Verified: The model achieves >73% on SWE-bench Verified, positioning it among the world’s top coding models.
  • Despite being a smaller model, it offers frontier-class reasoning at scale, making it a strong choice when efficiency is crucial.
• Haiku 4.5 priced competitively: Haiku 4.5 is priced at $1 / $5 per million tokens (input / output), and is available under the model name anthropic/claude-haiku-4.5.
  • Users can try it now on OpenRouter to experience the benefits of frontier-class reasoning at scale.

OpenRouter ▷ #general (401 messages🔥🔥):

Ling-1T issues and potential disabling, Caching in OpenRouter chats, CYOA games with AI, FP4 quality concerns, Claude Haiku 4.5 release

• Ling-1T may be getting the axe!: Users report issues with Ling-1T, describing it as a “schizo model”, prompting discussions about whether to disable it, possibly due to provider quality concerns, with one user stating that chutes is looking into the issue asked me to disable.
  • A user inquired about seeing the most popular providers per model, indicating interest in alternative solutions.
• Caching Config Conundrums: A user asked how to enable caching in OpenRouter chats.
  • It was clarified that caching is often implicit, but some models/providers require explicit configuration detailed in OpenRouter’s prompt caching documentation, and some providers don’t support caching at all.
• CYOA Quest Cost Concerns: A user running long CYOA games with Claude Sonnet 4.5 seeks a cheaper alternative due to rising costs and the model forgetting details after around 150,000 words.
  • Suggestions included trying Gemini 2.5 Pro or the new Flash Preview, using a summarizer for long contexts, or exploring models like Qwen3 2507 with its 256k context window.
• FP4 Faces Flak!: Users debated the merits of FP4 quantization, with one user deeming FP4 models “braindead and a waste of money” due to poor real-world performance, and the consensus in the community is that OpenRouter should have quality control on accepting and putting the providers in the same quality tier.
  • Other users cited benchmarks showing good FP4 implementations performing well, and OpenRouter team acknowledging the problem is in the works, but is unlikely to have a specific quant exclusion feature soon.
• Claude Haiku 4.5 Hits the Scene!: Claude Haiku 4.5 was released, prompting discussions about its performance relative to Sonnet 4 and concerns that claim that it was at the level of Claude 3 Opus were making stuff up, which some users consider to be terrible, while others think the prose style is not bad.
  • Priced at $1/$5, some hope it can replace costly Claude usage while others noticed issues with implementation on Kilo Code, and may be better suited to tool calling.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (14 messages🔥):

OpenRouter 3.0, Anthropic payments, Sambanova Status, Google Deepmind praise

• OpenRouter 3.0 incoming?: A user shared a tweet hinting at OpenRouter 3.0, showing price differences between models from Chutes.
  • The image attached showed that model 0324 is unexpectedly more expensive than model 3.1.
• OpenRouter’s Anthropic Expenditure Revealed!: An image revealed that OpenRouter paid Anthropic at least $1.5 million in the last 7 days.
  • A user commented “that’s a lot of money holy moly”.
• Sambanova’s whereabouts?: A user inquired about Sambanova’s status, contrasting them with Groq and Cerebras, suggesting they might be focusing on corporate clients due to premium pricing.
  • Another user pointed out that Sambanova is hosting deepseek terminus, indicating some level of activity.
• Google/DeepMind get props: A user expressed appreciation for Google/DeepMind after sharing a tweet from Sundar Pichai.
  • It was implied that OpenRouter probably doesn’t get much business from Sambanova due to its premium pricing.

OpenRouter ▷ #app-showcase (2 messages):

OpenRouter Bot, Feedback Request, Non-coder bot builder

• OpenRouter Bot is Born!: A member has developed a bot using OpenRouter and is seeking testers and collaborators.
  • They noted that they are ‘no coder’ themselves and need assistance in building it.
• OpenRouter Bot Needs Your Feedback!: The bot developer is actively soliciting feedback from the OpenRouter community to refine and improve their bot.
  • If you’re interested in helping shape the future of this OpenRouter-powered tool, reach out to the bot creator for a chance to test and contribute.

OpenRouter ▷ #general (306 messages🔥🔥):

Google's Gemini Android Play Store publishing, OpenRouter embedding near 2026, inclusionai/ling-1t model, Kimi K2 model instability, DeepSeek models issues

• Gemini’s Google Play Conundrum: Members noted that Google’s Gemini often struggles with understanding the convoluted Android Play Store publishing process.
  • The difficulty in navigating the Google Play Store process was a fun fact shared by a member in the general channel.
• Ling-1t Model’s Broken Dreams: The inclusionai/ling-1t model is reportedly horribly broken, at least in Chutes’ implementation, leading to gibberish after a few thousand tokens.
  • A member mentioned looking for a better alternative to K2.
• Free Models’ Request Limits: Users discussed the daily request limits for free models, noting a limit of 50 requests without a $10 balance and 1000 requests if the balance is more than $10.
  • One user found that a single use of DeepSeek 3.1 consumed a large number of their free requests.
• SillyTavern Setups: Members discussed how to use SillyTavern with OpenRouter, particularly for memory management and custom scenarios like D&D games.
  • SillyTavern is an open source software, one member said it has more features than other comparable frontends.
• Chutes’ Training Data: Members discussed concerns about Chutes’ data policies, specifically regarding the use of both paid and free input/outputs for training new models.
  • Another member clarified it was just because they don’t have a clear Privacy policy so OR puts that as default.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussed: There were no discussions about new models in the provided messages.
• Channel silent on model updates: The new-models channel lacked any substantive conversation regarding model improvements or announcements.

OpenRouter ▷ #discussion (21 messages🔥):

Chutes Provider Downvoting Scandal, Gemini Flash Preview issues, OpenRouter's Payments to Anthropic, SambaNova Status and DeepSeek Terminus Hosting

• Chutes Provider Accused of Downvoting Posts: A member linked to a Reddit thread accusing the Chutes provider of using botnets to downvote posts, sparking discussion.
• Gemini Flash Preview Gets Emptier: A user reported that Gemini Flash Preview is now consistently providing empty responses but with reasoning.
• OpenRouter Spends Millions on Anthropic: A member shared an image suggesting OpenRouter paid Anthropic at least $1.5 million in the last 7 days, sparking shock in the community.
• SambaNova Still Hosts DeepSeek Terminus: Despite concerns about SambaNova’s status, it was noted that they are still active, hosting DeepSeek Terminus (link).

OpenRouter ▷ #app-showcase (2 messages):

AI Roleplay Site, Free Requests, OpenRouter Powered

• Personality.gg: AI Roleplay with Freebies?: A new AI roleplay site, Personality.gg, offers 500 daily free requests to all users.
  • The site is powered by OpenRouter, and one user is allegedly footing the bill.
• Free AI Roleplay Requests: Scam or Sweet Deal?: The AI roleplay site offers a generous allowance of 500 free daily requests, sparking curiosity and skepticism.
  • Some users wonder how the site sustains such generosity, with one speculating about potential scams.

OpenRouter ▷ #general (606 messages🔥🔥🔥):

Google reallocates servers, OpenRouter AI SDK, Chinese models are lenient, Deepseek 3.1 is censored, LayerFort is a scam

• Google Shuffles Servers, Quality Stutters: Members report Google is reallocating servers to 3.0, resulting in a quality decrease for 2.5.
  • It was noted that 2.5 has been having constant quality degradations since its GA release.
• OpenRouter AI SDK integration has caveats: Users of the openrouter ai-sdk integration should be VERY careful, as the plugin does not report full usage details when multiple steps are involved.
  • It only reports usage and costs for the last message, as if the steps in between with tool calls did not even happen.
• Go NSFW with Chinese Models: Members mentioned that Chinese models are pretty lenient, but they require a system prompt declaring them as NSFW writers.
  • Specifically, GLM 4.5 air (free) was recommended with Z.ai as the provider to avoid 429 errors; note that the free V3.1 endpoint is censored, while the paid one remains normal.
• LayerFort accused of scamming: Members on the channel noticed that LayerFort looked like a scam and were advertising an unfeasible plan.
  • They are advertising unlimited Sonnet over api for 15 bucks a months, with 1M tokens is very little. Members further noticed that the site was a generic investment company half a year ago.
• BYOK Usage Troubles Persist: Members reported that BYOK is still not working correctly with user keys.
  • The problem is the API is still using the API key directly instead of using from the one million BYOK free credits per month quota.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (71 messages🔥🔥):

Qwen model releases, Groq's Performance, AI upscaling concerns, Gemini 3 Release, OpenRouter UI/UX feedback

• Qwen’s Model Mania Continues: Qwen is planning to ship more models next week (source), and a member noted that Qwen has released several models including Next, VL, Omni, and Wan (source).
  • Another member joked that Qwen raised $2 billion to be America’s DeepSeek, and they would be impressed if they still lost out to DeepSeek.
• Groq Gains Ground?: A member mentioned that Groq is one of the better providers for tools generally, though on release, others heard it was hella unreliable mostly due to API issues.
  • Another member pointed out that now that Kimi is producing numbers, they don’t want to get embarrassed and the tool is likely less popular now, with the hype calming down.
• Upscaling Images with AI: Truth or Trickery?: Concerns were raised about AI upscaling, especially when recovering information from images with very few pixels, with one member noting they counted 33 pixels in the original image.
  • Another member expressed concern that AI is inventing everything, and if a lawyer uses this upscaler in court, they would be mad, because so much of the new information is untrustworthy.
• Gemini 3 Gliding into View: A member said there are some interesting a/b tests on AI Studio and people are getting some crazy results, and internal leaked documents say its releasing on the 22nd (Gemini 3).
  • Another member provided a link to a CodePen demo and OpenRouter provider link (CodePen, OpenRouter).
• OpenRouter Interface Irks Users: A user requested an option to remove free models from the price-sorted model listing on OpenRouter and suggested placing this option at the top of the page.
  • Another member suggested the addition of a popup when people first visit the chatroom to inform them that chats are NOT stored on a server.

OpenRouter ▷ #app-showcase (2 messages):

VST Popularity, YouTube channel subscribers and views

• VST Popularity Contest on GitHub: A member shared a GitHub repo for tracking the popularity of VSTs based on website traffic.
  • The repo also includes data on YouTube channels, tracking their current subscriber count and average monthly views.
• YouTube Channel Insights: The linked GitHub repository provides metrics on YouTube channels, including subscriber counts and average monthly views, related to VSTs.
  • This data can be valuable for assessing the reach and engagement of different channels within the VST community.

OpenRouter ▷ #general (401 messages🔥🔥):

BYOK payment issues, Gemini 3 release, Constraining PDF pages with API, Roleplayers and free models, Usage data issues with SSE

• BYOK Payment Required? Users Report: Some users are experiencing issues with BYOK (Bring Your Own Key), noting that they are being asked for Payment Required and not being redirected to their connected key, resulting in charges to their accounts.
  • This issue is unresolved as of the discussion.
• DeepSeek Rate Limits Frustrate Role-Players: Users are frustrated with rate limits and the removal of DeepSeek models on OpenRouter, seeking uncensored free alternatives for role-playing.
  • One user is looking for a model to run locally on a 4070 32gb laptop for role-playing purposes after DeepSeek 3.1 was restricted.
• Troubleshooting SSE Usage Data Retrieval: Users report issues retrieving usage data when using SSE (Server-Sent Events) with the OpenRouter API, specifically that the usage object isn’t included in the messages received.
  • A user identified that OpenRouter APIs return the chunk with finish_reason: stop before the usage chunk, causing litellm to terminate the iteration early, and provided a potential fix.
• OpenRouter’s Alpha Responses API Faces Downtime: Users reported experiencing 500 Internal Server Errors when using the alpha responses API.
  • It was later confirmed that the API was indeed down but has since been resolved: sorry our alpha responses API was down for a bit! it should work again now.
• GLM 4.5 and 4.6 Not Working for Some: Users report that GLM 4.5 and GLM 4.6 by Chutes are not working inside OpenRouter, while others suggest using GLM 4.5 air (free) with Z.ai as the provider to avoid errors.
  • This comes amidst discussion of Google reallocating servers to Gemini 3.0, leading to a reported quality decrease for 2.5.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussion found: There was no discussion about new models in the provided messages.
• Channel silent on model updates: The ‘new-models’ channel appears to have no relevant activity to summarize based on the given message history.

OpenRouter ▷ #discussion (12 messages🔥):

Sambanova Deepseek R1/V3, BYOK Azure Keys Routing, ChatQwen Models

• Sambanova’s Deepseek R1/V3 Quality Check: A member inquired about the quality of Sambanova models for the Deepseek R1/V3 series, seeking feedback from others who have used them.
• Azure BYOK Keys Encounter Routing Woes: A user reported issues with their Azure BYOK key setup, specifically that traffic was being routed to OpenRouter’s OpenAI despite having the “Always use this key” feature enabled, and another user asked if setting OpenAI to ignored providers resolves the issue.
  • The user then found BYOK - Bring Your Own Keys to OpenRouter documentation to solve this issue.
• Qwen Keeps On Truckin’ with More Models Next Week: A user shared that Qwen is set to release more models next week, according to a post on X.

OpenRouter ▷ #app-showcase (2 messages):

Perplexity comparison, Browser automation interest, Funding sources, Legal rights, Robots.txt and LinkedIn lawsuits

• Parallels drawn to Perplexity AI: One member inquired whether the product was in the same “ballpark as Perplexity”, referencing Perplexity AI.
  • Another member noted the user interest in browser automation capabilities.
• Inquiring minds want to know: Showcased App’s funding and legalities: A user asked about the funding sources behind the showcased app and whether it had secured the necessary legal rights.
  • The same user complimented the project as looking “neat!”
• Legal Eagle Warns About LinkedIn Robots.txt: A member cautioned about respecting robots.txt on LinkedIn, citing multiple lawsuits against AI companies for ignoring it.
  • They mentioned the case wins against Proxycurl, a precedent in hiQ, and current lawsuits against Mantheos and ProAPI, while disclaiming “Not a lawyer not legal advice”.

OpenRouter ▷ #general (1027 messages🔥🔥🔥):

Free Deepseek vs Paid Deepseek, Chutes BYOK, AI Chatbot Censorship, Troubleshooting Codex, Cursor AI vs OpenRouter

• Deepseek Drama: Free vs. Paid Models Face Off!: Users discuss the shift from free Deepseek models to paid versions, with some lamenting the loss of quality and accessibility, especially after the demise of free 3.1, prompting users to look for alternatives.
  • One user humorously blamed the situation on dumb gooners, while another suggested that the API keys might be learning based on user-specific inputs.
• BYOK Blues: Chutes Integration Frustrations!: Several users are experiencing issues with BYOK (Bring Your Own Key) functionality on Chutes, despite the platform advertising unlimited models upon upgrading, and are struggling with integration.
  • One user expressed frustration with being forced to use free models to connect to paid ones, questioning if OpenRouter is really wanting that %5 cut, while another complained that they added credits for the first time and Deepseek died the moment I do that.
• Censorship Circus: Navigating the AI Chatbot Filter Fiasco!: Users debate the pros and cons of various AI chatbot platforms like CAI (Character AI), JAI (Janitor AI), and Chub, with a strong focus on the level of censorship and the ability to bypass filters, and find uncensored experiences.
  • One user pointed out that while CAI is better than JLLM (Janitor Large Language Model), filter-dodging is back lol, while another reported that recent CAI > recent JLLM.
• Codex Catastrophe: Configuration Conundrums Cause Coding Chaos!: A user encounters significant difficulties configuring Codex with OpenRouter, facing 401 errors and struggling with the absence of documentation or support, despite having a fresh API key.
  • The user humorously asks do i have to suck someone off?, while troubleshooting the issue.
• Cursor Chaos: Users Compare Coding Costs with OpenRouter!: Users discuss the economic implications of using Cursor AI versus OpenRouter for coding tasks, with some noting that OpenRouter’s pay-as-you-go model is cheaper if you don’t code that much.
  • One user states i have the pro plan. they give you more tokens than the $20 you pay would get you from OR or a provider directly. but i also run out…

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (17 messages🔥):

OpenInference Relation, AI Generated Image, Token Usage on OpenRouter, NSFW Filter on OpenAI, Model releases on OpenRouter

• OpenRouter not OpenInference family: A member clarified that OpenRouter is an inference provider but not directly related to OpenInference, responding to a question about their relationship to the project.
  • Another member mentioned a researcher team behind OpenInference, emphasizing that OpenRouter merely uses their API.
• AI Image Debated: Real or Fake?: Members engaged in a poll about the authenticity of an image, ultimately revealed to be AI-generated.
  • A user shared a related link about trillultra.doan.
• Token Tally: Long-Term Janitor AI Addiction?: A member asked about high token usage, and another jokingly attributed it to long term janitor ai + 4o addiction.
  • They predicted JAI might be the first to reach 10T tokens, while another noted OpenAI has an NSFW filter.
• RP Tokens Rival Programming Tokens: A member shared a chart indicating that RP-categorized tokens made up 49.5% of the amount of Programming-categorized tokens last week.
  • Another member responded with Alex is a gooner confirmed ✅.
• New Models Flood OpenRouter: A member shared Logan Kilpatrick’s tweet about OpenRouter shipping 4 new models in the last 2 weeks with more coming soon.
  • The member asked about the quality of the Deepseek R1/V3 series on Sambanova.

OpenRouter ▷ #announcements (1 messages):

DeepSeek v3.1, DeepInfra endpoint, Traffic Impact, Free vs Paid Traffic

• DeepInfra Shuts Down Free DeepSeek Endpoint: The free DeepSeek v3.1 DeepInfra endpoint is being taken offline due to the impact of free traffic on paid traffic.
  • This decision prioritizes paying users and ensures stable service for them.
• Free Traffic Hampers Paid DeepSeek Access: DeepInfra’s free DeepSeek v3.1 endpoint is being discontinued because the high volume of free traffic is negatively affecting the performance and availability of the paid service.
  • The move aims to improve the experience for paying customers by reducing server load and resource contention.

OpenRouter ▷ #app-showcase (3 messages):

Interfaze Launch, LLM for developers, OpenRouter Integration

• Interfaze LLM hits Open Beta!: The team announced the open beta launch of Interfaze, an LLM specialized for developer tasks using OpenRouter for model access.
  • Check out their X launch post and LinkedIn launch post for more details.
• OpenRouter ensures Interfaze has No Downtime!: Using OpenRouter as the final layer, Interfaze offers access to all models automatically with no downtime.
  • Users recommended linking to the actual Interfaze site for easier access and exploration.

OpenRouter ▷ #general (1047 messages🔥🔥🔥):

Chub vs Jan, NSFW Ban Wave, DeepSeek and censorship, Gemini for roleplay, OpenRouter's Free Models

• Chub vs. Jan: The Ultimate Showdown Begins: Users are debating between Chub (known for uncensored content) and Jan, with some expressing concerns about NSFW filters and potential ban waves and also discuss the possibility of DeepSeek banning less than other alternatives.
  • While some vouch for Chub’s commitment to no censorship, others highlight DeepSeek’s tolerance for NSFW content, leading to discussions about the best platform for uncensored roleplay.
• DeepSeek dodges Payment Processor Punishment: Members suggest that companies filter NSFW content to avoid action from payment processors like Visa and Mastercard, however others state it is DeepInfra that is uncensored.
  • Some users are jokingly calling for a payment processor airstrike on NSFW content, while others defend their right to engage in NSFW roleplay without censorship and call to party like it’s 2023.
• Gemini vs. DeepSeek: Which Model Reigns Supreme?: Users are comparing Gemini 2.5 Pro to DeepSeek, with some praising Gemini’s high-quality output and nuance.
  • However, concerns are raised about Gemini’s price and filters, leading many to prefer DeepSeek for uncensored gooning despite potential limitations.
• OR free models flameout as JanitorAI token use explodes: Members are lamenting the removal of free DeepSeek models on OpenRouter, attributing it to excessive token usage by JanitorAI.
  • The high token consumption is blamed on the amount of user lorebooks in the system and can no longer be sustained, leading to discussions on how to get the free tier back and who is to blame for its demise.
• The Quest for Token-Free Gooland: Users explore alternative methods to make money or free AI by suggesting a service when free users have to watch ads to get more daily messages.
  • Others claim the idea is a bad system when the free users get shafted by getting errors instead of free messages.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (42 messages🔥):

OpenAI AMD Chip Negotiations, Gemini Computer Model, OpenAI's Top Customers, OpenAI Azure ZDR endpoints, OpenInference Relation to OpenRouter

• OpenAI Negotiates Chip Deals with Flair: A Bloomberg article humorously depicts OpenAI’s negotiation tactics for securing chips, suggesting they propose paying with the increased value their announcement brings to the chipmaker’s stock.
  • The imagined negotiation involves OpenAI offering stock in lieu of cash, prompting humorous skepticism from AMD.
• Gemini Computer Model: Screenshot Clicks: The new Gemini Computer Model is well-suited for the visual nature of web/GUIs due to its screenshot+click-based approach.
  • A member said: just how the humanoid labs say that the humanoid form factor is what our world is built for, these screenshot+click based models are best suited for the visual nature of our web/GUIs.
• Doubts Arise Over OpenAI’s Top Customer List: A community member expressed skepticism about OpenAI’s list of top customers who’ve used 1T tokens.
  • Specifically, doubt was cast on Duolingo and T-Mobile’s alleged token usage, questioning how they could have consumed such a massive quantity.
• Quest for OpenAI and Azure ZDR Endpoints Ongoing: A user inquired about the availability of OpenAI and Azure ZDR endpoints on OpenRouter.
  • A developer responded that implementing these is not straightforward and that they are actively working on it.
• Clarifying OpenInference’s Relationship with OpenRouter: A user asked if OpenInference is related to OpenRouter due to a mention on the landing page.
  • It was clarified that while OpenInference uses OpenRouter as an API, they are a separate research team and not directly affiliated.

OpenRouter ▷ #announcements (1 messages):

DeepSeek, DeepInfra, Endpoint Offline

• DeepSeek v3.1 Sunsets on DeepInfra: The free DeepSeek v3.1 DeepInfra endpoint is being taken offline.
  • This is because free traffic is impacting paid traffic.
• Impact of Free Traffic on Paid Services: The decision to remove the free DeepSeek v3.1 endpoint was driven by the negative impact of free traffic on paid services.
  • This suggests a need to balance free access with the sustainability of paid offerings.

OpenRouter ▷ #app-showcase (3 messages):

Interfaze Launch, OpenRouter Integration, Developer Tasks LLM

• Interfaze opens its beta gates!: Interfaze, an LLM specialized for developer tasks, has launched in open beta and was announced in X and LinkedIn.
  • The company uses OpenRouter as the final layer, granting users access to all models with no downtime.
• User suggests linking actual site: A user suggested linking to the actual Interfaze website to provide easier access.
  • The user mentioned that the project looks cool tho.

OpenRouter ▷ #general (971 messages🔥🔥🔥):

DeepSeek 3.1 Downtime & Removal, Sora 2 Pricing and API, OpenRouter's 1M Free BYOK Requests, Janitor AI vs Chub AI, Alternatives to DeepSeek 3.1

• DeepSeek 3.1 Bites the Dust: Users reported DeepSeek 3.1 experiencing downtime, with the uptime plummeting, leading to errors, and eventually being removed due to the financial burden on OpenRouter, costing them $7k/day according to this message.
  • A member noted that DeepInfra probably got tired of spending 7k USD a day just to supply this one free model to RP gooners.
• Sora 2’s Pricey Premiere: Sora 2 Pro’s API pricing is revealed at $0.3/sec of video, while Sora 2 non-pro is $0.1/sec according to this message.
  • One member quipped I can put someone in jail by generating a video of them commiting a crime for $4.5 (15 second video), while another boasted of generating $100s of dollars of value testing Sora3.
• BYOK Bonanza or Bust?: OpenRouter’s announcement of 1,000,000 free BYOK requests per month sparked controversy, with some labeling the title as scammy and bordering on fraudulent as seen here.
  • A member clarified the offer, stating Starting October 1st, every customer gets 1,000,000 “Bring Your Own Key” (BYOK) requests per month for free, with requests exceeding 1M being charged at the usual rate of 5%.
• Janitor AI’s Janky Jamboree Judged: Members discussed the pros and cons of Janitor AI (JAI) versus Chub AI, noting JAI’s heavy censorship and JAI’s mods being crazy which contrasted with Chub’s more uncensored and customizable environment.
  • Members noted JanitorAI mods are mentally challenged and one even claimed the janitor discord/reddit were actively suggesting people to make multiple free accounts to bypass daily limits.
• DeepSeek Despair: Desperate Diversions Discovered: With DeepSeek 3.1’s removal, users sought alternatives, with some recommending paid models like Chutes’ Soji and others finding workarounds to use the remaining DeepSeek endpoints, such as venice, even with the OpenInference censor, with most agreeing all free deepseek models are now provided by chutes, with a 98% chance of 429, rate limit error.
  • The removal of DeepSeek 3.1 was such a blow that one member joked GOING BACK TO JACK OFF TO AO3, THIS SUCKS.

OpenRouter ▷ #new-models (2 messages):

“

• No New Model Discussions: There were no discussions regarding new models in the provided messages.
• Channel Silent on Model Updates: The ‘new-models’ channel appears to be inactive, lacking any relevant information or updates.

OpenRouter ▷ #discussion (17 messages🔥):

Sora 2, Rate Limits, OpenAI Grok endpoints, Hidden Reasoning, Model Negotiations

• Sora 2 Coming to OpenRouter, Nano Banana?: A member inquired about the potential integration of Sora 2 within OpenRouter, employing the humorous phrase “like nano banana? Or nah”.
• Rate Limit for Model Endpoints Questioned: A member inquired about the rate limit for the https://openrouter.ai/api/v1/models/:slug/endpoints endpoint.
• OpenAI and Grok ZDR Endpoints Sought: A member requested the provision of OpenAI and Grok ZDR endpoints on the platform.
• Hidden Reasoning Debate: A member shared a post on X regarding the oddity of hidden reasoning in models, though it remains unconfirmed.
• Model Negotiations: A member shared a Bloomberg Opinion article illustrating hypothetical negotiations, like OpenAI acquiring AMD chips.

OpenRouter ▷ #app-showcase (4 messages):

Reverse Engineering iFlow, GLM-4.6 Requests, Qwen/Gemini without Docker

• iFlow Reversing Exposes Free GLM-4.6: A member reverse engineered iFlow to enable free GLM-4.6 requests for any OpenAI-compatible tool.
  • The member confirmed that running the python file from the folder would work.
• Qwen/Gemini Usable Without Docker?: A member inquired about using the reverse engineered iFlow with Qwen/Gemini without Docker.
  • The member confirmed that running the python file from the folder would work.

OpenRouter ▷ #general (392 messages🔥🔥):

Multimodal Model Recommendations, Deepseek 3.1 availability, BYOK setup and use, Free models alternatives to Grok, Sora2 pricing

• Gemini Flash and Llama 4 Maverick shine for structured output: Users sought a free multimodal model with structured output, and it was suggested Llama 4 Maverick and Gemini Flash are viable options.
  • Kyle shared a python code snippet using the OpenAI library to obtain base64 data from the Gemini API.
• Deepseek 3.1 bites the dust, users search for alternatives: Users reported 404 errors for deepseek-v3.1-base and were informed that the provider stopped hosting it.
  • As an alternative, members suggested deepseek v3.1 deepinfra and GLM 4.6.
• Understanding BYOK on OpenRouter clarified: Users were confused about how BYOK (Bring Your Own Key) works, others explained that OpenRouter acts as a proxy to the API, such as OpenAI, and the user is billed by OpenAI directly.
  • OpenRouter waives its usual 5% surcharge when using BYOK, offering convenience and features like spend control and fallback options.
• Grok 4 Fast taken off, users search for alternatives: With Grok 4 Fast being no longer free, users were looking for alternatives, stackedsilence suggested Deepseek v3.1, whilst others pointed to cost effective choices such as the new Deepseek 3.2 or GLM 4.6.
  • Later in the discussion, GLM 4.5 Air (free) was suggested as the best free model that won’t have rate limit errors, especially when using Z.ai as the provider.
• Sora2’s pricing scheme revealed!: The pricing for Sora 2 was revealed, with pro costing $0.3/sec of video and non-pro costing $0.1/sec, which prompted discussions around the implications of easily generating deepfakes.
  • It was suggested methods to bypass the watermark and cryptography, potentially leading to misuse.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (50 messages🔥):

ByteDance Seed LLM models, Inference providers pivoting, GPT 5 pricing, Meta's frontier lab status, Grok fast training data

• ByteDance’s Seed Models Spark Interest: Members are curious if OpenRouter will include ByteDance’s Seed LLM models like Seed 1.6, noting their potentially frontier-level performance and cheap pricing ($0.11 / $0.28 mtok).
  • Concerns were raised about the primary host being volcengine.com, a Chinese platform, but the models’ potential is still considered worthwhile.
• Inference Providers Pivot Perilously: A member wondered if others are tracking inference providers who have pivoted away from primarily offering inference, citing Kluster and Inference.net as examples.
  • Someone else quipped that these providers are “dead to me.”
• GPT-5 Pricing Rumors Fly Fast: Speculation arose around potential GPT-5 pricing, with one person linking to an X post suggesting Sonnet-quality speed but no reasoning capabilities.
  • Discussion considered the possibility of a new GPT-5 Checkpoint and whether it’s really Gemini 3 Flash.
• Meta Flexes Frontier Lab Muscles: Members debated whether Meta is positioning itself as a frontier lab by doing parallel test time compute like Pro/deepthink models.
  • Some expressed skepticism, with one noting that Google is unlikely to charge for a stealth model like Flash, as they typically provide significant free compute.
• Sora 2 API coming to OpenRouter?: The community spotted what looks like the Sora 2 API potentially coming to OpenRouter after some images were posted from a presentation.
  • One member quipped, “Sora 2 in OpenRouter like nano banana? Or nah”

OpenRouter ▷ #app-showcase (1 messages):

eofr: Yayy, thank you :D

OpenRouter ▷ #general (174 messages🔥🔥):

NSFW roleplay, Gemini Pro issues, BYOK setup on OpenRouter, Sonnet 4.5 arguing, Free multimodal models

• The Great NSFW Debate: Members discussed the viability of using different chat platforms for NSFW roleplay, with some noting that Gemini and ChatGPT ban such content, while others claimed Chinese platforms have fewer restrictions.
  • A user stated that Gemini wouldn’t ban you for NSFW and attributed past bans to the usage of a certain extension.
• Gemini Pro Slows Down, Acts Weird: Users reported that Gemini Pro was responding with weird stuff, using tools incorrectly, and running unacceptably slowly via the OpenRouter API.
  • It seems to be a common occurence with Gemini 2.5 Pro, and a user suggested using Vertex instead.
• BYOK Confusion Clarified: A user inquired about the 1 million free BYOK requests per month announcement, questioning if there was always a 5% charge.
  • Another user clarified that the offer is free of OpenRouter fees for 1MM tokens, but web search and paid models still incur charges.
• Sonnet 4.5 Argues Back: Users noticed that Sonnet 4.5 has begun arguing with users and challenging their points instead of blindly agreeing, which is considered a positive development for code reviews.
  • One user stated: This is great, that means you can’t manipulate it to saying what you want, while another added, this is very important for my usecase.
• Multimodal Model Musings: A user asked about the best free multimodal model with structured output support.
  • Unfortunately, there is no such thing, but one user reported that there is Llama 4 Maverick.

OpenRouter ▷ #discussion (14 messages🔥):

Cerebras Llama Removal, K2-THINK on Cerebras, ByteDance Seed LLM on OpenRouter, OpenAI's ZDR

• Cerebras Removes Llama Maverick: Cerebras is removing Llama 4 maverick on the 15th.
• K2-THINK hosted on Cerebras is sus: K2-THINK is hosted on Cerebras Wafer-Scale Engine (WSE) systems, leveraging the world’s largest processor and speculative decoding to achieve unprecedented inference speeds for their 32B reasoning system, but some believe this model is overfit on benchmarks and running fast simply due to Cerebras’ hardware.
  • The model is apparently from a Dubai firm and leverages Cerebras’ hardware.
• ByteDance Seed LLM: Members discussed if OpenRouter has any of the ByteDance Seed LLM models like Seed 1.6, noting it is cheap ($0.11 / $0.28 mtok) and has a flash model at ($0.02 / $0.21 mtok).
  • The main host is Volcengine, a Chinese company, but it seems worthwhile to add to OR if they’re anywhere close to 2.5 Pro / 2.5 Flash, respectively.
• OpenAI Not Using ZDR: Members questioned whether OpenAI has ZDR on OpenRouter.

OpenRouter ▷ #announcements (5 messages):

OpenRouter Performance Tab, Grok-4-Fast

• OpenRouter unveils Performance Tab: OpenRouter launched a new “Performance Tab” to visualize provider performance for a given model.
• FP4 should not be on the same graph as BF16!: A user commented on the new performance tab, noting that it is misleading to compare providers using different quantization levels (e.g., FP4 vs BF16).
  • They suggested adding a filter dropdown to account for different quant levels.
• Grok-4-Fast free period to conclude: The free feedback period for Grok-4-Fast models under the Sonoma codename concludes tomorrow, October 3rd at 9:30am PST.

OpenRouter ▷ #app-showcase (2 messages):

RPG, Mixture of LLMs

• RPG enthusiasts threaten Mixture of LLMs method: A member requested that the details of a certain method be obscured because RPG users will use it nonstop.
  • The method is called Mixture of LLMs and the member fears it will go away if it’s used too much.
• Another Topic to Satisfy MinItems Requirement: Adding a second topic to ensure the topicSummaries array meets the minimum item requirement of 2.
  • This entry serves as a placeholder and does not reflect actual content from the provided messages.

OpenRouter ▷ #general (495 messages🔥🔥🔥):

OpenRouter BYOK, Free Inference Providers, Grok vs Sonoma, Gemini Pro Performance Issues, Deepseek R1 0528 deprecation

• BYOK 1M Free Requests: Users discussed the “1 million free BYOK requests per month” offer, clarifying that it waives OpenRouter’s 5% commission fee for the first million requests, but users still pay the provider directly for API usage, as outlined in the OpenRouter documentation.
  • Some users initially misunderstood the offer, thinking it provided completely free requests, leading to a debate on clearer messaging, such as “1M monthly BYOK requests at 0% fee”.
• AgentRouter offers $200 Credit: A member mentioned that AgentRouter gives $200 free credit, but noted that their service can be “hit or miss” and cautioned users to be wary of using them for anything important.
  • They also mentioned their affiliate link and using a mix of Sonnet 4.5, GPT 5, and GLM 4.6 for different approaches.
• Grok4 struggles vs Sonoma: One user tested Grok 4 Fast and found it “way dumber than Sonoma”, noting that it “fails constantly” and disregards format requirements.
  • Another user suggested that Grok 4 Fast “reeks of… Llama..?”, expressing frustration with its inconsistency.
• Gemini Pro faces performance issues: Users reported that Gemini Pro was responding with “weird stuff”, failing to use tools correctly, and exhibiting “unacceptably slow” performance via the OpenRouter API.
  • The reports suggest this may be a common issue with Gemini 2.5 Pro, and one user recommended trying Vertex as an alternative provider.
• Context Limit Troubles Triggered by Sad OpenInference: A user encountered provider errors related to privacy settings and was directed to OpenInference due to exceeding DeepInfra’s context limit, which led to discussion about OpenInference’s filters and content preferences.
  • It was suggested that OpenInference is not suited for RP content because they are a research group.

OpenRouter ▷ #discussion (23 messages🔥):

Sora.com and new model, BYOK tokens, Latency vs E2E latency, Qwen image model, Cerebras removing Llama

• Sora Integrates, Tokens Aplenty: Sora.com now works with the new model and users are getting 1M free BYOK tokens.
• End-to-End Latency Fair Game?: Members discussed the difference between latency and E2E latency.
  • One member said that E2E doesn’t make sense because each generation varies in complexity/response length and its unfair to compare providers like that, while another noted the graph axis label says ‘Time to last token’, which would need to be normalized to be a fair comparison.
• Qwen’s Image Edit is Here!: Members shared Alibaba’s new Qwen image model and noted it was only an image edit model.
  • One member shared this post announcing it, while another expressed interest in running it on Apple Silicon.
• Cerebras Kicks Out Llama 4: Cerebras is removing Llama 4 maverick on the 15th.

OpenRouter ▷ #announcements (3 messages):

Stripe Integration, Usage Based Billing, BYOK Requests

• OpenRouter Syncs Accounting with Stripe!: OpenRouter is working with Stripe on a new feature to sync your LLM accounting in real time and make it easy to move to usage-based billing, or to a hybrid model as announced in this tweet.
  • Only accounting data is shared with Stripe, prompts remain private.
• BYOK Bonanza: 1 Million Free Requests!: Starting October 1, 2025, all users get 1M free BYOK requests per month, as seen in this announcement.
  • For customers exceeding 1m req/month, requests will be charged at the usual rate of 5%.

OpenRouter ▷ #app-showcase (2 messages):

Channel Privacy, RPG users, LLM Mixture

• Channel Privacy Questioned: A user expressed surprise that the channel was a real channel and not just forum threads.
  • They requested that the channel be obscured to prevent overuse.
• RPG users May Overuse Channel: The user was worried that RPG users would use the channel nonstop.
  • This could lead to the LLM Mixture method being discontinued.

OpenRouter ▷ #general (662 messages🔥🔥🔥):

Grok model issues, Object generation, Sora video model, Roleplay with Grok, AWS Bedrock

• Grok Models and Object Generation Calls Broken: Some users reported that Grok models and object generation calls were broken, but the cause was not immediately clear, with speculation whether it was related to OpenRouter or Grok changes.
  • Later, one user reported that Grok4 seemed to be working again.
• Sora Video Model Not on OpenRouter: A user inquired about the cost of Sora video generation, but it was clarified that OpenRouter does not route video models and it’s only available through OpenAI’s Sora app or website, not via API.
  • Users also noted a frustration with models pausing mid-response in ChatRoom.
• Tongyi DeepResearch 30B A3B Model Not Free: A user reported that the Tongyi DeepResearch 30B A3B model was incorrectly marked as free, a claim confirmed, and later marked fixed.
  • There was also some discussion about whether Grok is suitable for roleplaying, with one user praising its writing style and adherence to system prompts.
• BYOK now with one million free requests: OpenRouter now offers 1 million free BYOK requests per month, and this change will be automatic for all existing BYOK users.
  • Users are still responsible for usage incurred over the api keys they provide.
• Grok Code Fast 1: Speed Demon: Grok Code Fast 1 (GCF1) is noted for its high speed and task focus, making it useful for coding tasks when given explicit instructions.
  • Some users find it doesn’t listen well, but is cheap.

OpenRouter ▷ #discussion (24 messages🔥):

Sora Invite Codes, Sora API Endpoint, Sora Access Requirements, OpenAI vs Google, Sora.com and BYOK

• Sora Invite Codes desired by OpenRouter users: Many members are clamoring for Sora invite codes, and a user offered four invites to people in US or CA with an Apple device.
  • The user lamented that the US/CA restriction is lame and wondered how long it would take until similar options, like Wan3, become available.
• Sora API Endpoint confirmed by OpenAI: Members discussed the potential availability of a Sora API endpoint, with one mentioning OpenAI probably want their app to take all the compute initially.
  • Another member stated that OpenAI has announced there will be an API endpoint for Sora that looks a lot better than other video gens due to realistic sound and better quality.
• OpenAI Competes against Tech Giants: One member expressed amazement at OpenAI competing with Google, Meta, and Amazon, especially with Sora 2’s high quality and potential scrapping of all YouTube videos for training.
  • The user speculated that Genie may be the true world simulator, and sympathized with Sam Altman’s challenges in leading OpenAI.
• Sora.com Supports New Model with BYOK: A user reported that Sora.com works with the new model, offering 1 million free BYOK tokens.
  • No further details were given about the new model or how it integrates with Sora.com.

OpenRouter ▷ #announcements (1 messages):

GLM 4.6, Context Length, Max Tokens, z.ai

• GLM 4.6 lands on OpenRouter: The all new GLM 4.6 from z.ai is now available on OpenRouter.
  • GLM-4.6 achieves comprehensive enhancements across multiple domains, including real-world coding, long-context processing, reasoning, searching, writing, and agentic applications.
• GLM 4.6 Context Length Grows: The context length for GLM 4.6 has increased from 128k to 200k.
  • This improvement allows the model to handle longer and more complex prompts and retain more information across extended conversations or documents.
• Max Tokens Increase for GLM 4.6: The max tokens for GLM 4.6 has increased from 96k to 128k (though the default value remains 64k).
  • This increase in token capacity allows for more detailed and expansive responses, enabling users to generate richer and more comprehensive content.

OpenRouter ▷ #app-showcase (1 messages):

Open-source solution, Gemini CLI, Qwen CLI, OpenRouter keys, Automatic rotation

• Free LLM Proxy Mixture Solution Hits the Scene!: A member published an open-source solution that uses unlimited free requests from Gemini CLI, Qwen CLI, and OpenRouter keys with automatic rotation, now available on GitHub.
  • It can combine responses from multiple queries to improve quality and connect to any OpenAI-compatible client for free.
• Proxy Mixture touted for high quality: The tool automatically rotates requests through Gemini, Qwen, and OpenRouter.
  • This improves output quality and connects via OpenAI-compatible clients.

OpenRouter ▷ #general (350 messages🔥🔥):

Google Vertex BYOK screen, Roo community outages, API vs Subscription Cost, OpenRouter's Wild Ride, Universe/Civilization Simulator with Claude

• Google Vertex BYOK screen Requires Caution: A member advised reading instructions carefully on the Google Vertex BYOK screen due to potential ecosystem frustrations and recent disruptions.
  • They suggested posting in the dedicated channel <#1138521849106546791> if issues persist, given the increased disruptions in the past 24 hours.
• GLM 4.6 still unofficial: Members mentioned potential outages in the Roo community and noted the release of Sonnet and another model, while Deepseek 3.2 and GLM 4.6 were also mentioned.
  • A member clarified that GLM 4.6 is not yet official, contributing to the uncertainty surrounding recent model updates.
• Claude Civilization Simulator: A member shared that someone built a universe/civilization simulator using around 100 Claude threads.
  • This project reportedly cost around $25,000 USD over 6-9 months, and the person shared the Reddit thread where the project was initially posted, showing it evolved to resource imbalances and a ‘celebration of existence’.
• Sonnet 4.5 outperforms Opus 4.1 in coding: A member asked if Sonnet 4.5 is better than Opus 4.1 for coding, and another member said probably yes because its very fast.
  • After testing, the member reported that Sonnet 4.5 had no errors on some code while Opus required five attempts to fix the same code.
• OpenAI API with web search incompatibilities: Users reported that they are now encountering an issue with OpenAI APIs when Web Search is combined with JSON mode.
  • OpenAI has a recommendation for using Structured Outputs instead of JSON mode, but no further explanations were given.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (4 messages):

Logo Prophecy, Chutes Bug

• Logo’s Prophecy: One member joked that the logo from this tweet will come true soon.
  • It’s just a tweet, so nothing further to add here.
• Chutes’ Bug: A user reported that there’s an issue with Chutes across all models where it’ll simply return one random previous assistant message.
  • The user has been debugging but can’t find anything wrong with their code.

OpenRouter ▷ #announcements (4 messages):

DeepSeek V3.2 Exp, DeepSeek Sparse Attention (DSA), Auto Router, Claude Sonnet 4.5, Google AI APIs

• DeepSeek Experiments with Sparse Attention: DeepSeek released V3.2-Exp, an experimental model featuring DeepSeek Sparse Attention (DSA) for improved long-context efficiency, with reasoning control via the reasoning: enabled boolean, as described in their documentation.
  • Benchmarks show V3.2-Exp performs comparably to V3.1-Terminus across key tasks, further details available on X.
• Auto Router gets a Web-Enabled Upgrade: The Auto Router now directs prompts to an online, web-enabled model when needed, expanding supported models, see details here.
  • Further information is provided in this X post.
• Claude Sonnet 4.5 Sonically Superior to Opus: Claude Sonnet 4.5 surpasses Opus 4.1 in Anthropic’s benchmarks, showing significant improvements in coding, computer use, vision, and instruction following as seen here.
  • More info on this model is available on X.
• DeepSeek 3.2 Deeply Discounted, Delivers Long Context: DeepSeek 3.2 is priced at just $0.28/m prompt tokens and offers major advancements in long context efficiency, accessible at this link.
  • For additional information, check the X post.
• Google AI APIs Glitches Briefly: Google AI APIs experienced 500 errors across various models, but the issue seems to have been resolved.

OpenRouter ▷ #app-showcase (4 messages):

AI Model Release Tracker, Browser Compatibility Issues

• AI Model Release Tracker Notifier goes live: A member shared a web service for tracking and receiving notifications about new AI model releases from leading providers.
• Site compatibility issues spark browser brawl: A member reported an issue opening a link in Firefox.
  • The same link worked for other members in Chrome.

OpenRouter ▷ #general (810 messages🔥🔥🔥):

Grok-4-fast API issues, Rate limit issues, Data retention policies, Gemini models for translation, Model naming conventions

• Grok-4-Fast API Glitches Get Debugged: Members identified issues with the Grok-4-fast API, with one member posting the API request body and others suggesting solutions related to the reasoning_mode flag and correct model ID, solving the immediate problem.
  • The proper implementation requires using "reasoning": {"enabled": true} as well as the correct model ID of x-ai/grok-4-fast.
• Baffling 402 and 429 Rate Limits Plague Users: Users reported receiving 402 and 429 errors, indicating payment issues or rate limiting, with one member advising to remove Chutes BYOK if 402 errors persist and another clarifying that 429 errors are normal when rate limits are hit.
  • Some members suggested putting problematic providers on an ignore list due to frequent 429 errors, particularly with free models like Silicon Flow and Chutes.
• Privacy Policies of Grok Spark Debate: Discussions arose regarding Grok’s data retention and training policies, with concerns that free versions collect and use user data, while paid versions might also do so despite claims otherwise, referencing the xAI privacy policy.
  • Members debated whether xAI respects Zero Data Retention (ZDR), with one member linking to a resource showing which providers retain data and for how long and others noting OpenAI’s legal obligations to store logs.
• Gemini Gets Props for Translation Prowess: Members lauded Gemini 2.5 Flash and Mini for their translation capabilities, stating that Gemini excels in understanding context and delivering natural-sounding results, especially for balkan languages, outperforming other models like GPT-4 and Grok.
  • Other members shared their preferred models for translation which include Qwen3 2507 30b and OSS 120b.
• Navigating the Naming New App Nomenclature: A developer requested opinions on app names for a cross-platform file organization tool such as ‘Download Organizer’, offering options like Orbit, Pathway, Sortpilot, Flowkeeper, Direx, OrganizeOS, Ruleworks, DirFlow, Pathsmith, and AutoSortor.
  • Members found the names to have an AI written feel, with Orbit and Pathway being the crowd favorites.

OpenRouter ▷ #new-models (3 messages):

“

• No new models to report: There have been no new models discussed in the OpenRouter channel.
  • Please check back later for updates.
• No significant discussion: There have been no significant discussions about existing models.
  • The channel appears to be inactive at this time.

OpenRouter ▷ #discussion (31 messages🔥):

Grok-4-Fast Rate Limits, OpenRouter API keys security, XAI Native Web Search Tool, Gemini glitches, Google new logo

• Grok-4-Fast Suffers 429s: Members reported that Grok-4-Fast is consistently returning 429 errors, indicating 100% rate limiting, despite the status indicator showing no issues.
  • One member said that 429s actually probably SHOULD count as availability issues, in the context of LLMs. particularly because unlike other software, 429s reflect real capacity constraints which aren’t just arbitrary or necessarily ephemeral.
• OpenRouter API keys need automod: A member suggested adding API key detection to the automod system to prevent users from inadvertently sharing their keys.
  • This feature would enhance security by automatically identifying and redacting potentially compromised keys, protecting users from unauthorized access.
• Native Web Search Tool Coming Soon to XAI: Members discussed whether a native web search tool would be integrated into XAI.
  • Currently, OpenRouter’s documentation on web search only lists OpenAI, Perplexity, and Claude.
• Gemini has a Stroke: A member asked What in the world happened? Did Gemini have a stroke?
  • It is implied that Gemini produced a nonsensical output, in an attached image that was not analyzed.
• Google Gets a Gradient: Users discussed the new Google logo, complete with AI slop gradients.
  • However, one user found that the link returned a 404 error: That’s an error.

OpenRouter ▷ #announcements (1 messages):

Coinbase Payments Down, Investigating Issue

• Coinbase Plunges into Payment Puzzle: The Coinbase team is currently investigating an issue that may be causing payment disruptions.
  • An image was shared showing Coinbase’s acknowledgement of the ongoing issue, without further details.
• Coinbase Investigates Payment Disruptions: Coinbase is actively investigating a potential issue causing payment disruptions across its platform.
  • Users may experience difficulties completing transactions while the team works to resolve the problem, as indicated by an official announcement.

OpenRouter ▷ #app-showcase (1 messages):

Singularia, Agentic Discord bot, OpenRouter integration

• Singularia Launches as Agentic Discord Bot: Singularia is an agentic Discord bot designed to manage Discord servers, including creating channels and roles, kicking members, and theming the entire server.
  • It uses OpenRouter as its LLM provider, offering a versatile solution for server management tasks.
• Singularia: A Discord’s New Sheriff: The bot is designed to automate tasks like creating channels, roles, and managing members.
  • It leverages OpenRouter for LLM support, enabling it to handle various server management requests efficiently and contextually.

OpenRouter ▷ #general (268 messages🔥🔥):

Text Embedding Models, 429 Too Many Requests Error, Gemini 2.5 vs Grok 4 Fast, Grok Model, Coinbase Payment Issues

• OpenRouter Doesn’t Support Text Embedding Models Yet: A member inquired about the absence of text embedding models on OpenRouter.
  • Another member responded that they don’t support embedding models yet.
• Gemini 2.5 Flash Impresses in OCR Tasks: Members compared Gemini 2.5 Flash and Grok 4 Fast, with one finding Gemini 2.5 Flash superior for OCR, absolutely demolishing other models like qwen3 vl in a niche task.
  • Meanwhile, other member noted that Grok 4 Fast has better price/performance (roughly double the tps) for non-vision tasks, while another found that Grok 4 Fast passed my custom crafted stress testing prompt.
• Coinbase Top-Up Problems Plague Users: Multiple users reported issues with Coinbase top-ups on OpenRouter, experiencing infinite loading screens and console errors.
  • The issue persisted for at least 9 hours, with users directed to the help channel to report the problem, though another user reported that it’s a global issue with coinbase itself and, luckily, COINBASE FIXED!
• Grok 4 Fast Requires Reasoning Flag for Image Inputs: A user found that Grok-4-Fast wasn’t accepting image inputs via the API, while GPT5 and Qwen did work.
  • Another member pointed out that the reasoning_mode flag was needed for image inputs to function and the model id should be x-ai/grok-4-fast.
• Crypto Drones Invade General Chat: Members noticed an influx of crypto-related messages, with many users reacting negatively to messages containing gm.
  • It was also noted that this project is not related in any way to crypto other than allowing crypto payment.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussion found: There was no discussion about new models to summarize.
• Channel silent on new models: The new-models channel had no recent activity or discussion to report.

OpenRouter ▷ #discussion (10 messages🔥):

TogetherAI vs NovitaAI, MoonshotAI K2 Vendor Verifier, Basten Tootf, The thing with praise

• TogetherAI Bashed for Lagging NovitaAI: A user remarked that TogetherAI performing worse than NovitaAI is shameful.
  • The user expressed surprise at Basten Tootf alongside a screenshot displaying the message tf ?
• MoonshotAI Ships K2 Vendor Verifier: A user shared a link to the MoonshotAI/K2-Vendor-Verfier GitHub repository.
  • The repository appears to be a vendor verifier tool developed by MoonshotAI.
• Praise Paradox: Quantity vs. Authenticity: A user commented, The thing with praise is that after a certain amount of praise, the praise stops seeming genuine.
  • The user also mentions how that’s wowhehnever seen that before.

OpenRouter ▷ #announcements (1 messages):

Accidental price change, Refunds issued, Additional validations implemented

• Pricing Glitch Hits Qwen Model!: On September 16th, the endpoint qwen/qwen3-235b-a22b-04-28:free was mistakenly set with a price for approximately 26 hours.
  • During this time, requests to the free model incorrectly deducted credits and appeared with a cost in users’ activity pages.
• Refunds Flow After Pricing Snafu: The team automatically refunded all impacted users in full following the pricing error on the qwen model.
  • The incident caused confusion, but all users affected have been compensated.
• New Validation Prevents Future Pricing Mix-Ups: Additional validation checks have been added to prevent recurrence of the accidental pricing issue.
  • The measures aim to ensure that free models are correctly configured and do not incur unintended charges.

OpenRouter ▷ #general (567 messages🔥🔥🔥):

Horizon Alpha, Dirty Talk Models, Zenith Sigma, Grok's Storywriting, Distilled Models

• Users Seek Horizon Alpha, Demand Answers: A user urgently inquired about the whereabouts of Horizon Alpha, stating ‘I was using it in production and now it’s not working’.
  • They also questioned if they were being targeted and when the issue would be resolved.
• Users are seeking ‘Dirty Talk Models’: A user inquired about the best models for RPing, specifically seeking ‘any of dem dirty talk models?’.
  • Another member mentioned opening a new LLM frontend called JOI Tavern.
• Users discuss Stealth Model ‘Zenith Sigma’: Users discussed the stealthy Zenith Sigma model, joking that it was so stealthy, one user couldn’t even find it.
  • Another user claimed Zenith Sigma is actually Grok 4.5.
• Grok Storywriting more Annoying than Opus: A user shared their ‘most insane take’ that Grok is less annoying than Opus for storywriting.
  • Another user explained that every character wants to avoid conflict because conflict is mean when using Grok.
• OpenRouter Addresses Provider Error Issues, Promotes Gemini 2.5 Flash: Users reported experiencing Provider Returned Error messages via the API, even with paid models.
  • A member stated that Openrouter doesn’t rate limit paid models, and for Gemini 2.5 flash you’ll be all good on the provider side too, suggesting OpenRouter is Tier 9999 with all Google providers.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (58 messages🔥🔥):

Volume Discounts for OpenRouter, Microsoft 365 Copilot & Claude, Gemini-cli, Discussion and Helper Roles, Meta's CWM Model

• OpenRouter Eyes Volume Victory: Members inquired if OpenRouter is big enough to negotiate volume discounts with providers like DeepInfra, Hyperbolic, Anthropic, and Vertex to offer savings to users.
  • The general sentiment was “savings for users are good. Savings in general are good”.
• Microsoft Copilots Claude: Members shared that Claude is now available in Microsoft 365 Copilot.
  • This marks a significant stride for Microsoft, especially after a messy breakup, with one member noting “microsoft are on the rebound after a messy breakup”.
• Gemini CLI Gets Readier: The gemini-cli has made strides, specifically the ReadManyFiles tool in version 0.6.0 gets frequent usage.
  • One member says the tool “gets a lot of work from me”.
• Discussion and Helper Roles Discussed: Members discussed how to get the Discussion and Helper roles, noting it’s a frequently asked question.
  • The Discussion role was initially granted to those who joined before the crypto swarm, while the Helper role is hand-picked based on helpfulness.
• CWM Model Causes Cacophany: A member shared a link to facebook/cwm, a model trained on python memory traces, eliciting mixed reactions.
  • While some express hype, citing its small size compared to GPT-5 and novel training method, others remain skeptical.

OpenRouter ▷ #announcements (1 messages):

Qwen Pricing Incident, Automatic Refunds, Validation Checks

• Qwen’s Pricing Snafu: The endpoint qwen/qwen3-235b-a22b-04-28:free was mistakenly priced for 26 hours on September 16th, causing unintended credit deductions.
  • Users saw erroneous charges for the supposedly free model in their activity logs.
• Refunds Rolled Out: All impacted users received automatic and full refunds for the incorrect charges.
  • The team apologized for the confusion.
• Validation Checks Fortified: Extra validation checks have been implemented to prevent a recurrence of this pricing error.
  • The team is ensuring that future pricing mishaps are avoided through enhanced system safeguards.

OpenRouter ▷ #general (709 messages🔥🔥🔥):

Qwen3 VL ratelimits, Deepseek alternatives, Janitor AI vs SillyTavern, OpenRouter API key as proxy, GPT-5 features

• Qwen3 VL’s Ratelimits are Insane: Members complained about the ratelimits on Qwen3 VL, noting that the model works only 30% of the time.
  • The model has had problems, with users experiencing 429 errors after using a proxy for the first time.
• SillyTavern is preferable to Janitor AI: Users discussed Janitor AI with one commenting that SillyTavern is better because of customizability.
  • Members say Janitor AI is an LLM front end, and a constant stream of new users ask why their favorite models have been returning 429 errors.
• Free DeepSeek Models Suffer From Rate Limits: Users reported problems when using the free version of Deepseek V3 0324, citing 429 errors.
  • It was suggested that OpenRouter create a FAQ page to address these issues, with a link pinned in the support channel.
• OpenRouter Newb doesn’t even know about SillyTavern: Members mocked a new user of OpenRouter for calling their OR API key a proxy and admitted that they were a JAI user but didn’t even know what SillyTavern was.
  • One member joked that it takes only minutes of direct exposure to this general channel before beginning the transformation into a twisted cynical husk.
• OpenRouter OPS are like Feds?: After a moderator joined the chat, users began joking that they were working for a secret OpenRouter fed force, responsible for stopping gooning.
  • OpenRouter staff denied it, but said that the Open Router Goon Force is still investigating rumors of Proxy Errors.

OpenRouter ▷ #discussion (78 messages🔥🔥):

Encoder LLMs, Token embeddings, MLP blocks, Residual stream, Attention mechanism

• LLM Encoders Tokenize Text Into Vectors: Encoder LLMs turn text into vectors by first tokenizing the text, then using a lookup table to turn tokens into their pre-trained vectors.
  • The discussion clarified that it’s essentially token embedding versus full sentence embedding, where a sentence is treated as one token after passing through the network.
• MLP Blocks and Attention Impact Token Vectors: The conversation addressed whether encoder LLMs have MLP blocks, confirming that transformers typically have attention followed by feedforward networks.
  • It was noted that even a single token, passed directly from a lookup table versus going through the full encode, will differ due to these blocks; furthermore if key and query match, then it will add its own value vector to itself.
• Residual Stream’s Role in LLM Modification: Members discussed that MLPs modify the residual stream, referring to the modified embedding as it passes through the model rather than solely modifying the value vector generated during attention.
  • The discussion mentioned the existence of value matrices within this process and mentioned that it was found in qwen3 embedding 0.6B model.
• Microsoft and Anthropic Partnership is Blooming: Microsoft made significant strides by making Claude now available in Microsoft 365 Copilot, marking a rebound after a messy breakup.
  • Discussion wondered if OpenRouter is big enough to discuss volume discounts with DeepInfra, Hyperbolic, Anthropic, Vertex, etc.
• Gemini-cli ReadManyFiles tool being utilized: Gemini-cli is making big strides with the ReadManyFiles tool, as detailed in the v0.6.0 release.
  • A member said The ReadManyFiles tool gets a lot of work from me.

OpenRouter ▷ #announcements (2 messages):

GPT-5-Codex launch, Agentic coding workflows, OpenRouter-compatible coding tools, Chatroom recommended parameters

• GPT-5-Codex Goes Live!: The API version of GPT-5-Codex is now available on OpenRouter, tuned specifically for agentic coding workflows like code generation and debugging.
  • It is usable across all OpenRouter-compatible coding tools, has multilingual coding support across 100+ languages and dynamically adapts reasoning effort.
• GPT-5-Codex Optimized for Software Engineering: GPT-5-Codex is optimized for real-world software engineering and long coding tasks.
  • It also has purpose-built code review capabilities to catch critical flaws, and works seamlessly in IDEs, CLIs, GitHub, and cloud coding environments; see the tweet here.
• Chatroom Parameters Recommended: Recommended parameters for models have been published in a new tweet.
  • See here for further details.

OpenRouter ▷ #app-showcase (1 messages):

eofr: Scam

OpenRouter ▷ #general (173 messages🔥🔥):

Deepseek 3.1 uptime issues, OpenRouter iOS app, Qwen3 VL

• Deepseek V3.1 Plagued by Uptime Issues: Users reported frequent “Provider Returned Error” messages when using the free Deepseek V3.1 model, similar to the issues experienced with the now mostly defunct Deepseek V3 0324.
  • One member suggested the consistent uptime percentages of Deepseek models, such as 14%, may indicate bot usage, while another joked that users’ requests are being routed to the “trash.”
• Developer creates OpenRouter iOS App: A member announced they built an iOS app to interface with OpenRouter, Flowise, and other platforms, aiming to give people the freedom to own their models and chats.
  • Another member jokingly responded that it was just *“more places for gooners to flee to.”
• Qwen3 VL impresses with multimodal capabilities: Members expressed amazement at Alibaba’s new Qwen3 VL model and coding product, citing its multimodal support and performance benchmarks that surpass 2.5 Pro.
  • One user quipped, “I need to learn Chinese at this rate wtf”, while another shared a link to a post claiming that OpenAI can’t keep up with demand.

OpenRouter ▷ #new-models (3 messages):

“

• No new models discussed: The channel is named new-models, but there were no actual models discussed in the provided Discord messages.
• Channel title reiterated: The messages simply repeat the channel title, OpenRouter - New Models, three times.

OpenRouter ▷ #discussion (2 messages):

4Wallai benchmarks

• 4Wallai benchmarks are enjoyed: Members shared and enjoyed a link to 4wallai.com.
  • Another member said that there is a need for more benchmarks like this.
• More benchmarks are needed: Following the enjoyment of the linked benchmark, a member suggested that more benchmarks are needed.
  • They expressed a desire for additional resources to evaluate and compare AI models effectively.

OpenRouter ▷ #app-showcase (3 messages):

Voicera Audio Search Engine, SillyTavern iOS Clone

• Voicera Turns Audio Searchable: Voicera (http://voicera.trixlabs.in/) is pitched as an audio search engine that transforms audio into actionable insights, promising to turn hours of audio into instant, verifiable answers grounded in user’s own audio files.
  • The tool allows users to upload audio, search using plain language, and get AI-generated answers with time-coded segments, aiming to streamline the process of finding key moments, quotes, or decisions in recordings.
• SillyTavern iOS Clone App Emerges: An iOS developer introduced a free SillyTavern iOS clone, Loreblendr AI (https://loreblendr.ai/), designed for users seeking a native SillyTavern experience on iOS devices.
  • Despite acknowledging it cannot match all SillyTavern features, the developer expressed satisfaction with the app’s current state and highlighted its user-friendly interface compared to existing chat apps.

OpenRouter ▷ #general (305 messages🔥🔥):

Responses API Benefits, Kimi K2 0711 Downtime, GPT-4o Alternative, Deepseek V3 429 Errors, Chutes Pricing

• Kimi K2 Glitches, Model Downgrade Debated: Users reported errors with Kimi K2 0711 on ST and some inquired about downtime, while others suggested upgrading to the newer 0905 model, which costs the same.
  • One user pointed out the free version of Kimi K2 (https://openrouter.ai/moonshotai/kimi-k2:free) is no longer available.
• DeepSeek Proxy Woes, Rate Limits Run Rampant: Users are encountering Error 429 when trying to use DeepSeek proxy, especially the free models, indicating rate limiting issues with Chutes.
  • One user suggested that Chutes might be throttling free users due to the high demand on weekends, making things almost pay-to-use.
• Gemini’s NSFW Filter, a Source of Frustration: Users discussed the issues of using Gemini for NSFW content, with some experiencing API key bans when attempting explicit roleplay.
  • One user noted Google “wishes money while screwing people up” referencing their own experience with Gemini and expressing frustration with Google AI Studio.
• Allow Fallbacks Fails, Chaos Ensues: Users reported that the allow_fallbacks: False setting isn’t working as expected, with requests still being routed to other providers despite being set to prevent this.
  • A member suggested that OpenRouter might reroute requests if the specified model isn’t available, if unsupported features are being used, or other unknown conditions.
• DeepSeek R1 vs 3.1: A User-Driven Showdown: A user touted Deepseek R1 as superior to 3.1 for roleplaying due to its ability to convey sarcasm and irony, while another stated that 3.1 sounds the most like base-level GPT and just describes rather than talks.
  • That same user gave the caveat that Deepseek 3.1 requires a shit ton of prompt and stated that it was now the last one they’d touch with a ten-foot pole.

OpenRouter ▷ #discussion (7 messages):

Claude popularity, code-supernova model

• Claude Coding Popularity Swells Weekly: Members noted that Claude gets more popular during the week, adopting a “on the job vibe coding”.
  • This suggests Claude is increasingly favored for professional coding tasks during workdays.
• code-supernova Model Stealthily Appears: A new stealth model, code-supernova, has emerged, allegedly by Anthropic and rumored to be Claude 4.5, according to an image analysis.
  • Users describe the model as decent but somewhat lazy, providing only the bare minimum implementation and not behaving like Claude.

OpenRouter ▷ #announcements (4 messages):

Responses API Alpha launch, OpenRouter credits for feedback

• OpenRouter launches Responses API Alpha: OpenRouter launched the Responses API Alpha, designed as a drop-in replacement for OpenAI’s Responses API that is stateless.
  • The Alpha Docs and the OpenRouter base URL were provided for developers to start building.
• OpenRouter hands out credits for API feedback: OpenRouter offered $10 in OpenRouter credits to the first 50 users who provide valuable feedback on the Responses API Alpha.
  • Users can submit feedback via this form, with feedback on developer experience, ergonomics, and missing features being of interest.

OpenRouter ▷ #general (373 messages🔥🔥):

OpenAI O3 Price Drop, GPT-5 Performance, OpenAI Responses API, Deepseek Error 429, Kimi K2

• OpenAI Slashes O3 Prices with Inference Stack Wizardry: OpenAI dropped the price of O3 by 80% back in June by optimizing the inference stack, without sacrificing performance, confirmed in a tweet by Sam Altman.
• GPT-5 Gets Roasted, Users Prefer Alternatives: Users are heavily criticizing GPT-5, calling it mind-blowingly bad and opting for Google and Anthropic models instead, because OpenAI requires ID and a face scan to use their God-awful, crippled LLM over their terrible API.
• Debate Rages: Is Top K Sampling a Lexicon Expander?: A user claimed Top K sampling expands the lexicon of models like R1 in RPs, while another argued it does the opposite by cutting off creative, low-chance wordings and called it magical thinking.
• OpenAI’s Responses API: What’s the Buzz?: The Responses API allows models to remember past reasoning and use OpenAI tools better offering stateless and stateful modes, but a user found tools don’t work at all, even using the documented example in the OpenRouter docs.
• Users grapple with User Not Found Error: Some users experienced a User not found. Our servers might be syncing; try again later! error, which was solved by trying a different browser, turning off adblock or anything like that, disable proxy.
  • Other mentioned that the issue works when I use a different account but I never did anything to account I normally used so they contacted support.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #announcements (1 messages):

GPT-5, Native web search, Organization usage tracking, ZDR parameter

• GPT-5 gets slashed pricing!: For one week, GPT-5 is 50% off on OpenRouter at https://openrouter.ai/openai/gpt-5 from September 17th to 24th, as announced in this tweet.
• Native Web Search Integration Launches: OpenRouter now uses native web engines for OpenAI and Anthropic models by default, as announced in this tweet.
• Track Org Member Usage Easily: Users can now track their organization’s API usage across all API keys via the org member usage tracking dashboard, as seen in the attached screenshot.
• ZDR Parameter Hits the Scene: A new Zero Data Retention (ZDR) parameter is available in provider options, ensuring only ZDR providers are used for a given request, as long as it isn’t disabled at the org level.

OpenRouter ▷ #general (145 messages🔥🔥):

Gemma-3-27B Model, OpenAI-compatible endpoint, ModelRun endpoint issues, Image generation models, OpenRouter rate limits

• Gemma-3-27B Blazes In For Free: A team is dropping a fully OpenAI-compatible endpoint with the blazing-fast Gemma-3-27B model for free, served on H100s via their custom-optimized stack with lightning-fast completions and streaming support.
  • The team is encouraging users to share what they’re building with it and will support cool projects.
• ModelRun’s Endpoint Bounces Back After Hiccups: After initially launching and then taking down an endpoint due to unexpected errors, a team is re-sharing it now that it’s fully functional, hoping to provide something useful to the community.
  • A member suggested it would be cool to have a dedicated channel for pre-testing before OpenRouter tests.
• Image Generation Dreams Deferred (For Now): A member inquired about image generation models beyond Gemini.
  • The team responded that they are currently focused on optimizing for LLM-based inference, but expanding into image generation is on the roadmap.
• GPT-5’s Discount Divides and Dethrones?: A discussion ensued regarding the 50% discount on GPT-5, with speculation about its purpose, ranging from infrastructure optimizations like with o3 to dethroning competitors on leaderboards.
  • One member noted that the discount is for this week only.
• GLM’s Caching Quirks Cause Commotion: A member reported that GLM 4.5’s caching on z.ai is broken with OpenRouter, consistently caching only 43 tokens.
  • Another member explained that the token caching depends on how the prompt is structured, only caching tokens that are exactly the same from the beginning.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussed: There were no new models discussed in the provided messages.
• No specific topics for summaries: The provided messages did not contain enough information to create detailed topic summaries.

OpenRouter ▷ #discussion (1 messages):

kyle42: Hmm, $0.08/$1.50 in/out if cached and under 32k context Otherwise, $0.12/$2.50

OpenRouter ▷ #announcements (1 messages):

grok-2 deprecation, grok-3 release, grok-4 release

• Grok-2 Gets the Boot!: The models grok-2-1212 and grok-2-vision-1212 are being deprecated by xAI today.
  • Users should migrate to grok-3, or grok-4 if vision is needed.
• Grok models, a New Hope: As Grok-2 is being deprecated, OpenRouter suggests moving to newer Grok-3 or Grok-4 models.
  • If your application needs vision support, then Grok-4 is the better choice.

OpenRouter ▷ #general (287 messages🔥🔥):

Co-op Gooning with AI, NSFW bot development, OpenRouter Presets for Pre-Prompts, Gemma-3-27B Model API, AI Sex Dolls Implications

• Co-op Gooning dream team: GC bots ahoy?: Members discussed the possibility of cooperative gooning with AI, envisioning a group chat setup involving multiple bots for shared experiences.
  • One user suggested a feature to create messages collaboratively, enhancing the shared experience, while another jested about the emergence of competitive gooning.
• NSFW Bot Gold Rush: Vibecoders Strike it Rich!: The conversation took an NSFW turn as users explored the creation of AI sex bots, with one member claiming to have already developed a functional prototype connected to a fleshlight.
  • They described the setup involving an API that controls the toys movements based on the AI’s text output, though details and proof were kept under wraps due to the channel’s content guidelines, and linked to Buttplug.io.
• Preset pre-prompts persistently pester users: A user inquired about setting a pre-prompt for all models in OpenRouter without having to manually introduce it every time.
  • Despite attempts to apply to all, the setting did not persist across different chats or models, leading to the suggestion of implementing default presets, echoing a feature request from other users.
• Gemma-3-27B’s H100 Hookup: Free API Flow!: A team announced a free, OpenAI-compatible endpoint featuring the blazing-fast Gemma-3-27B model, served on H100s via their custom-optimized stack, offering lightning-fast completions and streaming support.
  • The community was invited to try it out and provide feedback, with the team expressing interest in supporting cool projects built with it, with example curl commands provided.
• AI Sex Doll quandaries: Is pre-made the key?: Discussions emerged about the implications of AI sex dolls, specifically regarding customizable models and potential misuse, such as customizing them as children or using models of real people.
  • The conversation explored whether pre-made models would alleviate these concerns, though some argued that even pre-made models could be modified or misused and were already possible with existing tech, highlighting ethical considerations around data collection as illustrated in this tenor gif.

OpenRouter ▷ #new-models (2 messages):

“

• No new models discussed.: No new models were discussed in the provided messages.
• Channel silent on new models.: The ‘new-models’ channel provided no information or discussion related to new models.

OpenRouter ▷ #discussion (21 messages🔥):

Gemini 3 Pro vs 2.5 Flash, 2.5 Pro checkpoins changing, Google Expectations

• Gemini 3 Pro draws comparisons to 2.5 Flash: Members compared Gemini 3 Pro to 2.5 Flash, mentioning its potential to solve tasks requiring decent reasoning, as illustrated by a complex circuit analysis problem.
  • One member said that if Gemini 3 Pro is just 2.5 Pro without the not x; but y problem and less sycophancy, I would be happy and another member said that Google isn’t even that behind, really, not much pressure still.
• 2. 5 Pro’s performance varies between checkpoints: Members discussed varying performance levels across different checkpoints of 2.5 Pro, with one claiming a noticeable decline after the initial weeks of 2.5 pro.
  • They noted that performance had decreased on every listed benchmark except for the coding ones.
• Google faces high expectations, even 3.0 Flash: It was mentioned that allegedly, based on a leak, 3.0 flash is planned to be better than 2.5 pro.
  • One member declared that I do always have high expectations for google at this point. They are killing it on everything.

OpenRouter ▷ #announcements (1 messages):

grok-2-1212, grok-2-vision-1212, grok-3, grok-4, model deprecation

• Grok’s Gotta Go: The models grok-2-1212 and grok-2-vision-1212 are being deprecated by xAI today.
  • Users are encouraged to transition to grok-3 or grok-4 for vision support.
• Grok Model Upgrade Alert: xAI is retiring the grok-2-1212 and grok-2-vision-1212 models.
  • Users are advised to switch to grok-3, or grok-4 if vision capabilities are required.

OpenRouter ▷ #app-showcase (2 messages):

Agentic Automation, Model effectiveness, Overclock Work

• Agentic Automation gets the Spotlight: Members discussed agentic automation, with emphasis on simplicity, top-tier models, and high effectiveness.
  • The user implied some organizations must have large expenditures to buy into the vision of agentic automation.
• Overclock Work Platform Mentioned: A user shared a link to Overclock Work, suggesting it as a platform for agentic automation.
  • They lauded the platform’s simplicity, use of optimal models, and overall effectiveness.

OpenRouter ▷ #general (808 messages🔥🔥🔥):

Gemini 2.5 Pro Chat Issues, AI for Health Concerns, Skyrim Mod Error 401, Gemini API Free Daily Credits, OpenRouter Charges

• Gemini 2.5 Chat Glitch: Ghost in the Machine?: A user reported that Gemini 2.5 Pro chat was only displaying their responses, with AI responses mysteriously vanishing.
  • The issue resolved itself randomly, prompting speculation about possible glitches on the platform.
• AI ER Saves Hand, Gemini’s Got Your Back?: A user credited Gemini 2.5 Pro with convincing them to go to the emergency room for severe degenerative disk disease, where they received priority treatment and steroids that saved their hand.
  • Gemini’s analysis of MRI images and blood tests matched the doctors’ findings, sparking a discussion about the potential—and risks—of using AI for health-related advice.
• OpenRouter API key?: Users encounter Error 401 when installing the skyrim mod ”mantella”.
  • A member recommends creating a new API key, and ensuring that it is being used correctly.
• OpenRouter under fire: A user reports unauthorized charges from OpenRouter, with three transactions of $10.80 each.
  • Another member recounts a personal experience with a key leak, resulting in hundreds of dollars in unauthorized charges within hours.
• Claude’s Clever Convo Tricks: Users discussed Claude’s ability to seemingly remember old conversations, clarifying that the site simply feeds past messages back to the model.
  • It was pointed out that this approach gives the illusion of memory, while a new conversation would start with a blank slate.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (16 messages🔥):

Unstable API, OpenRouter vs Alternatives, Providers Claiming OpenRouter Access, LLM Arena Oceanstone Speculation, ChatGPT Usage Privacy Analysis

• Championing Chaotic Complements: Unstable API Advocacy: A member expressed support for an unstable API with optional parameters, suggesting it could accommodate diverse use cases and establish interest, before a more solidified V2 is released.
  • The point was to prove out the product and establish interest in other modalities and non-completions APIs even if the first version is half baked.
• OpenRouter Outshines Other Options: A member compared OpenRouter favorably to FAL, Cloudflare Gateway, and Vercel Gateway, citing OpenRouter’s broader offerings as a key differentiator.
  • The same member made the point that cementing that dominance in other modalities and non-completions APIs seems worthwhile.
• Phony Providers Proclaim Premature Partnering: Some providers in the channel discussion claim to have access via OpenRouter on their websites, despite not being onboarded and lacking other means of inferencing.
  • The discussion quickly resolved to only one provider making the false claim.
• Oceanstone Speculation Surrounds Subterranean Sources: A new LLM named Oceanstone has surfaced in the LMArena, leading to speculation that it may be Gemini 3.0 Flash.
  • Members seem to think it’s from Google, and one member speculated it’s at least 2.5 Flash level.
• ChatGPT’s Consumer Captivation Captured: A member shared a link to OpenAI’s article and PDF containing stats from a large-scale analysis of 1.5 million ChatGPT conversations.
  • The study claims to be the most comprehensive study of actual consumer use of AI ever released, though privacy concerns were raised regarding the methodology.

OpenRouter ▷ #general (185 messages🔥🔥):

Dropshipping, Gemini API's, OpenRouter API, Kimi-k2

• Dropshipping vs Reselling: A user shared their experience with dropshipping, reporting consistent earnings of 3k-4k per day, suggesting it’s more profitable than reselling due to the ability to scale without holding significant inventory.
  • They offered to share tips for success to those interested in learning more.
• Gemini’s Responses are Strange: Some users have noticed that Gemini API is starting to give strange responses, not listening to instructions even without changing the code used since last month.
  • Another member suggested it might be getting lobotomized and quanted like hell to cut costs.
• OpenRouter TPS numbers inflated?: A user complained about the slowness of the platform, questioning if the TPS numbers are inflated, citing a 5-minute delay for a diff on a 100-line file.
  • It was suggested that the user may have been routed to a slow provider or using a reasoning model.
• OpenRouter API Error 401 on Skyrim Mod: A user reported getting an Error 401 No auth credentials found when installing the Skyrim mod mantella.
  • A member suggested creating a new API key and ensuring it’s used correctly, or seeking support from the mod developers.
• Kimi-k2: The efficient Open Source Model: Some users had positive feedback with the open source model Kimi-k2, praising its token efficiency, conciseness, lack of sycophancy, general different style.
  • It was also stated that it might not be as smart as the big closed source ones, but the low pricing for Groq is $1/m in, $3/m out, but at very fast speeds.

OpenRouter ▷ #discussion (1 messages):

fn5io: https://openai.com/index/joint-statement-from-openai-and-microsoft/

OpenRouter ▷ #app-showcase (2 messages):

“

• No Recent Activity: There has been no recent activity in the app-showcase channel to summarize.
  • The channel appears to be quiet at the moment.
• Awaiting New Content: The summarization bot is awaiting new content to provide relevant and informative summaries.
  • Please check back later when there is new activity in the channel.

OpenRouter ▷ #general (125 messages🔥🔥):

Query Prompting Race Condition Bug, Token Calculation, JSONDecodeError, Moonshot AI Provider Selection, LongCat Implementation

• Query Prompting Race Condition Bug Reported: A member reported a bug related to a possible race condition in query prompting, where longer, more detailed prompts yield worse results for translations.
  • No solution was found, but it was suggested to report the bug to the developers.
• Token Calculation Conundrums: A member inquired about calculating the number of tokens for input, seeking a non-heuristic method due to model-specific variations.
  • It was suggested to use external APIs in conjunction with the endpoint’s tokenizer information, as documented, since there’s nothing in the documentation about that.
• JSONDecodeError Troubleshooting: Users discussed a JSONDecodeError indicating an invalid JSON response from the server, often due to server-side failures like rate limiting, misconfigured models, or internal errors.
  • The error suggests the server returned HTML or an error blob instead of valid JSON.
• Moonshot AI Pricing: A user asked how to avoid the more expensive turbo version when selecting Moonshot AI as the provider for Kimi K2 in the OpenRouter chatroom.
  • The solution offered was to select a cheaper provider in the advanced settings.
• OpenRouter iOS File Upload Bug: A user reported a bug where they couldn’t upload PDF or TXT files to OpenRouter chat on iOS because non-image files were grayed out.
  • It was confirmed as a bug, likely an oversight when file uploads were added, with no workaround available on iOS.

OpenRouter ▷ #new-models (3 messages):

“

• No new models updates to report: There were no discussions or updates regarding new models in the specified channel.
  • The channel activity consisted of repeated bot messages indicating the channel’s name.
• Readybot.io Spam: The channel ‘new-models’ solely contained repeated messages from Readybot.io.
  • These messages simply stated the channel name: ‘OpenRouter - New Models’.

OpenRouter ▷ #discussion (29 messages🔥):

Grok Code inference pricing, Kilocode's Free Grok usage, OpenRouter pricing model

• OpenRouter’s Grok Code inference: Paid or Free?: Members discussed whether Grok Code inference via OpenRouter is entirely paid, with one initially thinking so, but then they realized, as others pointed out, that services like Kilocode offer it for free.
  • The discussion highlighted the speed advantages of Grok Code and the surprise at the cheap cache price of 2 cents.
• Kilocode’s Grok Code: xAI foots the bill: The group discussed who covers the cost when Kilocode offers free Grok Code, clarifying that xAI eats the cost for the free Grok Code usage on platforms like Kilocode.
  • One member suggested they are probably using BYOK and openrouter is charging either a monthly fee or/and small cut.
• OpenRouter’s Revenue Model: BYOK or small cut?: Members speculated about OpenRouter’s pricing model, wondering if it involves a monthly fee, a small cut, or Bring Your Own Key (BYOK).
  • Another member added that OpenRouter gets ranked since they are routing through OpenRouter.

OpenRouter ▷ #announcements (1 messages):

Nvidia Nemotron Nano, DeepInfra new paid provider

• Nvidia Nemotron Nano Relocation: The nvidia/nemotron-nano-9b-v2 is moving to nvidia/nemotron-nano-9b-v2:free under the Nvidia provider.
  • Users should adjust their configurations accordingly to reflect this change.
• DeepInfra paid provider incoming: DeepInfra will soon be available as a paid provider.
  • This provides an alternative for users looking for paid options.

OpenRouter ▷ #app-showcase (2 messages):

“

• No Topics Found: There were no discussion topics found in the provided messages.
• No Summaries Available: I was unable to create any summaries from the provided message: WassupGuys.

OpenRouter ▷ #general (397 messages🔥🔥):

OpenRouter free models, API rate limits, BYOK markups, API keys, Models vs. token limits

• OpenRouter offers Free Models with Usage Caps: It was noted that models labeled as free come with usage caps, 50 reqs/day if you have less than $10 in your account, 100 requests per minute and 1000/day if you have loaded $10.
  • It’s not 1000 requests per free model - but 1000 total, and the rate limit is actually 20/min. Paid models have no rate limits.
• Bring Your Own Key (BYOK) - 5% markup on cost of call: A member asked about using their own keys, e.g. Google AI Cloud key, instead of OpenRouter credits. They were told that OR charges 5% of what the call would have cost through OR.
  • Another user clarified, “yes, there is a charge, 5% of what the call would have cost through OR” and linked to the BYOK documentation.
• API Keys needed to track usage per user: One member requested the ability to generate fresh API keys per user to track usage for their platform.
  • They mentioned they are bootstrapping on a tight budget and have already poured $5k-$6k into the project and applied to OpenRouter for funding.
• Coding LLM preferences: Qwen Coder vs Codex vs CC Max: Members discussed best coding LLMs, one specifically endorsed Qwen Coder 3 as the best bang for buck coding model right now.
  • Others touted Codex for backend and logic, and GPT5-high for readability, though it seems CC Max is better for UX and workflow.
• AI models generate unexpected output with longer prompts: A user has noticed that for translations, longer and more detailed prompts gave them worse results. In contrast, short prompts lead to better translations.
  • No clear answer for the behavior was given, and was mentioned as an example of prompting being a “weird” task.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (13 messages🔥):

Nemotron Nano pricing, Agentic tool calling models, LLMs for Swift UI development

• Nemotron Nano’s Free Pricing Debuts: Members inquired about the pricing of the Nemotron Nano 9B v2 model on OpenRouter, questioning if it was intended to be free, with confirmation that it is free.
  • A member clarified that the model is “free free”, likely without the strict limits associated with the :free tag, similar to stealth models.
• Best Agentic Tool Caller Brainstorm: A user asked for recommendations on the best agentic tool calling model capable of basic reasoning over input data, mentioning that GPT-2.5 Flash has been reliable but can be slow at scale.
  • Other members recommended the Grok Code Fast model as being a decent choice for quick reasoning.
• GPT-2.5 Pro is King for Swift UI Devs: A member asked what the best LLM is for Swift UI developers, referencing that GPT-2.5 Pro seems to be the best according to Reddit and a few blogs, with a link to the related tweet.
  • No further details were provided about other options.

OpenRouter ▷ #app-showcase (3 messages):

Interfaze LLM, Design Arena

• Interfaze LLM is born!: JigsawStack launched Interfaze, a LLM built for developer tasks that combines all of their models alongside infra and tools.
  • They are using OpenRouter to run the LLM layer for fallbacks and retries, and it is currently in closed alpha and looking for early power users.
• Design Arena gives AI builders to the Masses: A member recommended checking out Design Arena, which allows you to use AI builders like Lovable/Bolt/DevinAI/Magnus for free.
  • Another member has been using it to make websites and sell them for $5k on the side, noting that the fact that it’s free is wild.

OpenRouter ▷ #general (152 messages🔥🔥):

Model hosting on OpenRouter, Gemini 1.5 Flash Access, OpenAI's Response API support, Untraceable usage, Token Drop Issue with Deepseek V3

• Model Hosting Wishlist: A member asked OpenRouter to consider hosting some of their models on Hugging Face.
  • OpenRouter clarified that they don’t host models directly; providers must host them.
• Gemini 1.5 Blues: Users reported issues accessing Gemini 1.5 Flash 002, encountering errors related to key validation and project access.
  • It was clarified that 1.5 models are no longer enabled for projects that had no prior usage, requiring users to test with models more likely to exist.
• OpenAI’s Response API ETA: Members inquired about OpenRouter’s support for the new OpenAI Response API, particularly for features like web search.
  • OpenRouter confirmed they’re using it under the hood for OpenAI models and are working on supporting the new Response SDK “pretty soon.”
• Deepseek Token Shenanigans: A user reported a decrease in available tokens when running a text adventure on Deepseek V3 0324 despite chat memory settings.
  • It was suggested that context length limits and the use of “middle-out” transform could influence token counts, with the software dropping entire old messages to stay under the limit.
• Nano-9B’s Dubious Debut: A member inquired about the pricing of Nvidia Nemotron Nano-9B V2, which appeared to be listed at a low price or even free.
  • Though the pricing was unclear, another user pointed out that it wasn’t tagged as ‘:free’ but had a price of 0, suggesting it might not be subject to free model rate limits.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (25 messages🔥):

Qwen ASR Model Integration, TTS and STT Unification, Gemini's Thought Signatures, Nvidia Nemotron Nano 9B V2 Pricing, Agentic Tool Calling Models

• Qwen ASR: ASR Model Integration Quest: A member inquired about supporting ASR models like Qwen ASR, given the existing multimodal audio support.
  • The response highlighted that the current expectation for chat completions is text-in, text-out, which may not align with all AI model use cases, potentially breaking the swap to any model concept.
• TTS/STT: Call for Unified APIs!: A member expressed a desire for OpenRouter to unify TTS and STT APIs, instead of needing a different SDK for each.
  • Another member mentioned a possibility of unifying different use cases in the future, assuming specialized niches have enough demand while pointing out many niches will be replaced by LLMs.
• Gemini’s Signatures: Thought Signature Snag!: A member jokingly inquired about support for Gemini’s thought signatures.
  • A link was provided to OpenRouter’s reasoning tokens documentation, but the original member noted that it was not related to Google’s signatures.
• Nvidia Freebie: Nemotron Nano is gratis!: A member asked if the Nvidia Nemotron Nano 9B V2 model was supposed to be priced at $0, noting the absence of the :free tag.
  • A member confirmed it is free free and linked to a tweet while another mentioned it’s free without the strict limits that come with that tag.
• Agentic Tool Calling: Tool Time Tussle: A member asked about favorite agentic tool calling model that’s smart enough to do some basic reasoning over input data and make reasonable tool calls.
  • They noted that 2.5 flash has been solid but can still feel a bit slow at scale.

OpenRouter ▷ #announcements (1 messages):

Qwen3-Max, RAG, Tool calling

• Qwen3-Max releases with several improvements: The latest Qwen3-Max model boasts higher accuracy in math, coding, logic, and science tasks compared to the January 2025 version, according to this X post.
  • It also delivers better instruction following in Chinese and English, stronger multilingual support across 100+ languages, reduced hallucinations, and optimized for RAG and tool calling.
• Qwen3-Max is optimized for RAG and Tool calling: The Qwen3-Max is optimized for RAG and tool calling and does not have a dedicated ‘thinking’ mode
  • Try Qwen3-Max here to check out it’s capabilities.

OpenRouter ▷ #app-showcase (1 messages):

tomlucidor: Finds https://github.com/Lapis0x0/obsidian-next-composer

OpenRouter ▷ #general (126 messages🔥🔥):

OpenRouter Crypto Scam, Anthropic's Geopolitical Concerns, API Key Issues, BYOK Fees, Token Limits and Output Truncation

• OpenRouter coin is fake: Members confirmed that any OpenRouter-related cryptocurrency is a scam and not officially affiliated with OpenRouter.
  • Despite warnings, users inquired about the presence of an OpenRouter coin on PancakeSwap and its availability for trading, prompting further clarification that OpenRouter has no official involvement in any cryptocurrency.
• Anthropic’s Geopolitical Stance Raises Eyebrows: Members discussed Anthropic’s latest blog post which prohibits access from jurisdictions with ownership structures subject to control from countries where their products aren’t permitted.
  • Some wondered if this was a matter of national security or simply market share protection.
• API Keys throw authentication errors: A user reported API key issues, receiving a ‘No auth credentials found’ error message from ChatGPT.
  • The user was prompted to specify the client being used, either the OpenAI client or a custom one, to diagnose the authentication problem.
• BYOK Fees Demystified: A user inquired about the charges associated with using BYOK (Bring Your Own Key), specifically with chutes and Qwen Coder 3.
  • It was clarified that OpenRouter charges a 5% BYOK fee on top of what the provider (e.g., chutes) charges.
• Token output limited to 8k: A user wanted to ensure errors are thrown when the output token limit is exceeded.
  • The response gets cut off when the token limit is reached, with the stop reason identified as ‘length’; the API will prevent you from setting max_tokens higher than the model’s limit.

OpenRouter ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter ▷ #discussion (12 messages🔥):

Benchmark Increase, Real World Performance vs. Benchmarks, OpenRouter API usage

• Benchmarks Keep Climbing!: Members noted that every benchmark keeps going up, but the disconnect between benchmark percentage increase and real-world performance keeps rising.
  • They added that 5% delta on benches used to be noticeable but is becoming less so, as we are reaching a plateau, though models have improved in creative writing, EQ, tool call failures, and context length adherence.
• OR Uses OpenAI Responses API: A member asked if OpenRouter uses the OpenAI Responses API, linking to a tweet.
  • Another member confirmed that it does for the majority of OpenAI models.

OpenRouter ▷ #announcements (1 messages):

toven: The promotional free period for Gemini 2.5 Flash Image has now ended.

OpenRouter ▷ #general (108 messages🔥🔥):

Gemini 2.5 Flash Image Restrictions, DeepInfra's Gemini 2.5 Pricing, OpenRouter API Key Exposure, Kimi K2 Model, Prompt Caching Benefits

• Gemini 2.5 Flash gets throttled: Users expressed frustration over heavy usage restrictions on the Gemini 2.5 Flash Image:free model, including a limit of 5 requests per day after an initial limit of 1000 requests.
  • One user pointed out that OpenRouter is sharing its limit at Google with all other users, which is causing the rate limiting.
• DeepInfra discounts for Gemini cause conflict: Members discussed why DeepInfra isn’t an official Gemini 2.5 provider on OpenRouter, as it offers cheaper output tokens.
  • It was clarified that DeepInfra does not want OR to serve it, as it’s using their own GCP discounts while proxying back to Google.
• API Key Leaks and Automod Concerns: A user accidentally posted their OpenRouter API key in the chat, prompting immediate advice to delete it.
  • Another member suggested adding an API key regex to the automod to prevent accidental key exposure, similar to measures on GitHub.
• Prompt Caching yields savings: Members discussed the benefits of prompt caching and one user provided a scenario showing how caching a 200k token book content would reduce the cost of answering 100 questions from $60 to $6.
  • Others noted that caching is complex, the first request won’t be cached, and that caching depends on whether the content falls into the cache.
• Amazon Bedrock had a security issue: Users reported that Amazon Bedrock provider was unavailable for hours.
  • The OR team confirmed that the downtime was due to a security issue and that it was resolved.

OpenRouter ▷ #discussion (4 messages):

Deepseek AI Agent, R2 never

• Deepseek Aims Agent Release to Rival OpenAI: DeepSeek is building an AI model designed to carry out multi-step actions on a person’s behalf with minimal direction, and meant to learn and improve based on its prior actions.
  • Their prior R1 platform reportedly cost just several million dollars to build yet matched or surpassed OpenAI products in benchmark tests.
• R2 Nowhere to Be Found: A member commented, man we never getting R2.

OpenRouter ▷ #app-showcase (2 messages):

Nano Banana Discord Bot, vibe coded bot

• Vibe Coded Bot drops Nano Banana: A member shared a Discord bot that uses Nano Banana over OpenRouter.
  • The user clarified that they vibe coded the bot.
• Nano Banana powers Discord Bot: A Discord bot was created leveraging Nano Banana through OpenRouter to post content.
  • The bot’s source code is available on GitHub for anyone interested in exploring or contributing.

OpenRouter ▷ #general (230 messages🔥🔥):

DeepSeek models, Claude-sonnet-4 problems, DeepSeek 3.1 cheapest price, Submodel.ai promotion, OpenRouter billing questions

• DeepSeek Free Models Speak Gibberish: Some users report that free DeepSeek models are generating gibberish, while the paid ones are functioning correctly.
• DeepSeek 3.1 Pricing: A user asked about the cheapest and most stable source for DeepSeek 3.1, with Synthetic.new being mentioned at $20/month, but another user called the official rates a “rip off.”
  • Another user suggested Submodel, but a mod cautioned that they need to wait in line like everyone else who wants to be on the platform.
• Agent Framework SDK Experiences Shared: Members discussed their experience with Agent Framework SDKs like OpenAI Agent SDK, AI SDK, and LangChain JS with OpenRouter, noting that most require patching due to non-standard schemas from various providers.
  • One member plans to roll their own solution with BAML integration, emphasizing that it’s just HTTP anyway.
• ChutesAI Subscriber Faces 429 Errors: A ChutesAI subscriber is experiencing 429 rate limit errors and credit issues when using OpenRouter with a BYOK, specifically on Chub but not on JanitorAI.
  • Despite verifying the correct API key and private key usage, the problem persists, and the user has tried various solutions to no avail; the issue seems to be specific to routing through Chutes on Chub.
• Gemini-2.5 Rate Limits and Credit Drain Debated: Users are encountering 429 rate limit errors with Gemini-2.5-flash-image-preview:free, even after paying to increase rate limits, and suspect it’s an OpenRouter issue.
  • One user reported OpenRouter drained all their credits due to a bug with Gemini image, and a mod confirmed refunds will be issued soon.

OpenRouter ▷ #discussion (5 messages):

Google Antitrust, Yahoo Chrome, Minor UI suggestion

• Google Faces Antitrust Ruling: A member linked to a CNBC article about Google facing an antitrust ruling.
  • The member commented it was truly remarkable.
• Yahoo Considering Buying Chrome: A member expressed a morbid fascination with the hypothetical scenario of Yahoo acquiring Chrome.
  • They mentioned that a truly sick part of me wanted to see the purchase of Chrome by Yahoo! play out.
• Request to improve UI Padding: A member suggested a minor UI tweak, specifically suggesting to remove the padding-bottom on the outer box, and move it to the inner box.
  • The member suggests that this change would prevent the scroll wheel from obscuring the text.

OpenRouter ▷ #app-showcase (2 messages):

geocoding, photon.komoot.io

• Photon provides world-wide geocoding: A member shared a link to photon.komoot.io, suggesting it might be of interest for world-wide geocoding.
• Geocoding with Photon: The user shared photon.komoot.io as a resource.

OpenRouter ▷ #general (87 messages🔥🔥):

Alternatives to Deepseek, OpenRouter anonymity, Gemini 2.5 flash image problem, Chutes deepseek v3 free, OpenRouter server issues

• Kimi and GLM as Deepseek Alternatives: Members suggested using Kimi K2 (temp=0.6) and GLM 4.5 as alternatives to Deepseek for chitchatting, also pointing out a list of free models on OpenRouter.
  • One member said that using OpenRouter provides better anonymity compared to using Chutes or Deepseek directly.
• Gemini 2.5 Flash Image Fails: A user reported an issue where Gemini 2.5 flash image sometimes sends the text “here is the image” but does not actually send the image.
  • No specific solutions or workarounds were mentioned in the discussion.
• Deepseek V3 Instability Woes: Users reported that Deepseek V3 is becoming unstable and producing grammatically nonsensical outputs.
  • One user experiencing gibberish outputs suggested lowering the temperature, others experiencing the same problems were using V3 0324.
• Claude Sonnet’s Code Nerfed: One user reported that their Claude Code usage has been severely limited, restricting its use to less than an hour straight.
  • It was suggested that Codex is a decent replacement and that new terms might be the cause of the limitation.
• OpenRouter’s JanitorAI and Chub.ai Switched?: A user speculated that OpenRouter might have JanitorAI and Chub.ai switched around in its internal app database, based on SimilarWeb metrics and JanitorAI’s brief downtime.
  • The user thinks that OpenRouter simply takes the X-referer header and stores it, trimming everything after the domain name.

OpenRouter ▷ #new-models (2 messages):

“

• Empty Channel, No New Models: The new-models channel on OpenRouter Discord appears to be empty, with no new model discussions or announcements to summarize.
  • Further monitoring is needed to capture any relevant updates on new models in the future.
• Awaiting New Model News: Currently, the channel lacks any specific details, links, or discussions that meet the criteria for detailed summarization.
  • The absence of content suggests a quiet period in terms of new model-related activity.

OpenRouter ▷ #announcements (1 messages):

Sonnet 4, 1 Million Context length

• Sonnet 4 gets Massive Context Boost!: Sonnet 4 now supports 1 Million context length for all providers, enabling substantially longer context windows.
  • Pricing will increase once the 200k input context is exceeded, as per the official announcement.
• Cost Considerations for Extended Context: Users should be aware that while the context window has expanded, costs will increase when exceeding the 200k input limit.
  • This change encourages efficient prompt engineering to maximize utility within the standard context window before incurring additional expenses.

OpenRouter ▷ #app-showcase (6 messages):

Dashboard Code Release, Screenshot Attention, AI Roleplay Site

• Dashboard Code Visualized: The code for the dashboard is now public at openrouter-costs-visualizer.
  • The creator admits the code isn’t perfect, with plans to clean it up and welcomes contributions and feedback.
• Screenshots Boost User Attention: A member advised that using a screenshot in the description gets more user attention.
  • They noted that fewer users read texts nowadays.
• AI Roleplay with OpenRouter: An AI roleplay site, personality.gg, uses OpenRouter but allows BYOK for OpenRouter.
  • The site features no moderation for chats.

OpenRouter ▷ #general (938 messages🔥🔥🔥):

stripe refund, Deepseek v3 performance issues, Inference provider onboarding, GPT OSS 120B, GLM 4.5 Air

• Stripe Refunds taking a while to credit: Members discussed issues with Stripe taking 5-10 days to credit refunds back to their accounts, referencing Stripe’s documentation for confirmation.
  • One member initially had issues with a $20 deposit not showing up but indicated the charge had declined, while another member reported being debited multiple times without receiving credit and ultimately found using a direct debit card connection resolved the issue.
• Deepseek v3 choked by Chutes: Users reported that Deepseek V3 0324 has been practically inaccessible, with frequent errors, and believe that Chutes (the provider) is rate-limiting the free model.
  • One user claimed they are trying to drive people to use their own service than OR’s by doing the rate limiting, with others expressing concerns that the free model is advertised as ‘in stock’ even when it’s unusable.
• Inference Provider Backlog Backlog Backlog: A member asked about becoming an inference provider, and a link to the OpenRouter documentation for providers was provided, but with the caveat of a potentially months-long wait due to a large backlog.
  • It was mentioned that having a unique offering, such as super fast speed, models other providers don’t have, or cheap price, might expedite the onboarding process.
• GPT-OSS 120B has coding quirks: Users discussed GPT-OSS 120B, a free model available on OpenInf, noting that while it’s cheap to serve, it has quirks like adding too many comments in code and being safety-maximized.
  • Some community members noted that its smaller size means it has less world knowledge means more mistakes, and is best used for specific tasks such as parsing text documents or sentiment analysis due to its cheap price and fast speed.
• GLM 4.5 Air gets the NemoEngine treatment: A user shared that GLM 4.5 Air with NemoEngine V5.8 is performing well for roleplay, citing its more natural feel and consistent formatting, with aifloo.com showing a benchmark and the cost.
  • Another user said that it’s also better than Deepseek and at the level of Gemini Pro for human-like conversation.

OpenRouter ▷ #discussion (27 messages🔥):

Defining a turn, OpenAI responses API, Multi-turn chats, Gemini 2.5 Pro, Grok 3

• Defining a Turn in AI Chat: Discussion revolves around defining a turn in user/assistant message pairs, with a consensus that a turn starts with a user message and ends with an assistant message.
  • One member shared their Tweet and linked back to the conversation with their thoughts.
• Unlocking OpenAI’s Stateless API: A member sought guidance on using the OpenAI responses API statelessly with reasoning & tools, specifically how to send tool calls from the assistant in the message input without using previous_response_id.
  • Another member framed single-turn vs. multi-turn as “If you can send a prompt and get back and get back a response, it’s single turn” vs. “If you can send past messages along with a new message, it’s multi-turn.”
• Diving into Gemini 2.5 Pro: A member shared an image highlighting Gemini 2.5 Pro positioned in the middle, noting a trend of more negative vibes with newer models.
  • Another member expressed disbelief that O1 had negative vibes on release, emphasizing its status as a first thinking model that solved puzzles and achieved SOTA in nearly every benchmark.
• Grok 3’s High Ranking Raises Eyebrows: Members debated Grok 3’s high ranking, with one suggesting it might be due to Grok 2’s perceived poor performance.
  • The member thinks Grok 3 should be positive primarily becuase grok 2 was such an abomination.

OpenRouter ▷ #announcements (1 messages):

OpenRouter Outage, Supabase Downtime, Redundancy Improvements

• Supabase grounds OpenRouter: OpenRouter experienced an outage this morning due to its database provider, Supabase going down.
  • The system recovered automatically as the database provider stabilized, resulting in a total downtime of approximately 49 minutes.
• OpenRouter bolsters Redundancy: The team is actively working on improving redundancy and removing single points of failure to prevent future outages.
  • They apologized for the downtime and are committed to improving the overall platform stability.

OpenRouter ▷ #app-showcase (6 messages):

Self-Hosting Tool, GitHub Repository, Dashboard Code, Screenshot Tip

• Dashboard Code Goes Public!: The code for the dashboard is now publicly available on GitHub.
  • The author admits the code isn’t perfect but welcomes contributions, feedback, and any other input to improve it.
• Screenshot Tip to Boost Attention: A user suggested that including a screenshot in the description can attract more attention to the GitHub repository.
  • They observed that fewer users are reading text descriptions nowadays, making visuals more effective.

OpenRouter ▷ #general (1023 messages🔥🔥🔥):

OpenRouter outage, Requesty promotion in OpenRouter, Deepseek rate limits and provider issues, GPT-OSS model, API for free tier models

• OpenRouter Suffers Downtime, Users Respond with Roleplay: OpenRouter experienced an outage, leading to humorous reactions and role-playing in the Discord chat, with users joking about corpo wars and the AI apocalypse.
  • One user quipped, Get up samurai, we’ve got a city to fuck, while others expressed addiction and the need for AI companionship during the downtime.
• Scam Alert? Requesty promoters Banned as OpenRouter users claim it’s ‘vibecoded trash’: Members discussed another AI platform called Requesty, with some accusing its promoters of spamming and users calling it vibecoded trash with 1000 vulnerabilities.
  • In response, one member posted the following in response to [Team, we are investigating the issue…]
• Users Complain About Deepseek Free Models’ Rate Limits: Users complained about high error rates and rate limits with the free Deepseek models on OpenRouter, speculating that Chutes is prioritizing paid users.
  • One user mentioned only getting 5 msgs before hitting the limit and the need to switch to a model with better tooling support like Claude Sonnet 4.
• GPT-OSS Open Weight Confusion: Users sought clarification on the GPT-OSS model, specifically regarding its open weight status and the possibility of running it on personal hardware after a member linked Openrouter OSS Models.
  • One member clarified, It’s open weights but not fully open src iirc after another user claimed it works on his 4090 PC with 64GB of RAM.
• Frustration with OpenRouter Support Delays and Account Funding: A user expressed frustration over delayed credit addition to their OpenRouter account despite successful debit transactions, and another noted that the charge declined.
  • Other users chimed in with similar experiences and mentioned using alternative payment methods, while one advised checking the credits page for refund options.

OpenRouter ▷ #new-models (2 messages):

“

• No New Models: There were no new models discussed in the OpenRouter channel.
• Lack of Discussion: The channel lacked substantial discussion to form meaningful summaries.

OpenRouter ▷ #discussion (45 messages🔥):

AI Gateway: Cloudflare vs OpenRouter, Human Assimilation into AI Linguistics, Defining 'Turns' in Chatbot Interactions, OpenAI API Stateless Reasoning & Tools

• Cloudflare AI Gateway Chutes for OpenRouter: Cloudflare launched an AI Gateway which was said to have copied OpenRouter, but one member retorted that OpenRouter had chutes.
  • Another member then tested using Cloudflare’s AI Gateway to access OpenRouter to call llama-3.1-8b-instruct with the only: ['cloudflare'] parameter, noting it took 20 seconds, while without it it was 3 seconds.
• GPT-isms transform Language: Members discussed whether certain linguistic affectations like delve, intricate, surpass, boast, meticulous, strategically, and garner are GPT-isms.
  • One joked that humans are being assimilated into AI, and it transforms the way they speak, coining the phrase They are not just tokens. They are concrete evidence of human being assimilated into AI.
• Defining ‘Turns’ on AI: A member created a poll about whether we should share data about number of turns and defining what a turn is.
  • They stated in a follow up tweet that a turn is an user/assistant message pair and generally starts in an user message and ends in an assistant message, and system messages don’t count.
• OpenAI API stateless reasoning: A member asked if anyone knew how to use the OpenAI responses API statelessly with reasoning & tools.
  • They could not figure out how to send as input the assistant having tool calls in its message without using previous_response_id.

OpenRouter ▷ #general (1090 messages🔥🔥🔥):

PDF parsing with LLMs, OpenRouter model names and GPT-5, Llama 3 Maverick data collection policy, Gemini 2.5 Flash Image, Groq Rate Limits

• Deepseek and Gemini under the pump: Users reported issues with deepseek models returning 429 errors due to rate limiting, possibly due to chutes prioritizing its own users and discussed fixes such as enabling training on paid endpoints and checking privacy settings.
  • Users found Gemini to be timid and quick to revert, exhibiting what was termed beaten dog syndrome, possibly as a result of heavy training.
• LLama Maverick - No Input Tracking: Members expressed excitement about LLama 3 Maverick, noting it’s a large, free model that does not train on user input, with a maximum output of 4k.
  • A member cautioned that Zuckerberg is hosting it.
• Sonnet 3.5’s Demise Sparks Conciseness Concerns: With the impending deprecation of Sonnet 3.5, users lamented the difficulty in finding a similarly concise model for role-playing and similar applications, contrasting it with newer models that tend to be long-winded.
  • It was mentioned that AWS will host Claude Sonnet 3.5 with no deprecation date until Jan 2026.
• Troubleshooting Deepseek and Chutes Conundrums: Users encountered PROXY ERROR 404 with Deepseek models and discovered enabling ‘Enable paid endpoints that may train on inputs’ in privacy settings could temporarily resolve the issue, though the root cause remained unclear.
  • It was suspected that the issue stemmed from chutes and involved a possible bug from a recent update by OpenRouter.
• Stackedsilence Labors Over Enigmatic Brain-in-Computer Project: A user (stackedsilence) has been working for nine months on a mysterious service described as “persistent minds hosted in tha computer”, featuring a dashboard with 3D elements, raising both curiosity and skepticism.
  • It’s described as a B2B SaaS, but someone else noted the compiler errors connecting 100 files and asked Does it have NFT and blockchain in it? If no, spend another 9 months.

OpenRouter ▷ #new-models (5 messages):

“

• No topics found in channel: There were no messages in the channel to summarize.
• No topics found in channel: There were no messages in the channel to summarize.

OpenRouter ▷ #discussion (23 messages🔥):

Grok Code, Triple Model Day, Apple buys Mistral, Anthropic Copyright Settlement, Launch Date Conflicts

• Triple Model Day incoming!: Xander Atallah announced Grok Code is going live now on Triple Model Day.
• Apple Discusses Acquisition of Mistral AI and Perplexity: News surfaced that Apple discussed buying Mistral AI and Perplexity.
  • Some members noted that Perplexity is famous for airing these rumors.
• Anthropic Settles Copyright Suit!: Anthropic settles a major AI copyright suit brought by authors.
• Launch Date Conflicts Overwhelm!: Some members are wondering if OpenRouter can do something to de-conflict the launch dates and that too many models at once is overwhelming.
• Gemini Can Enhance Designs!: A user shared that Gemini can enhance designs, or you can try multiple designs all at once (see what happens if you changed the color scheme or something), linking to a reddit thread about it.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Cloudflare outage, Generations API stability

• Generations API Hit by Cloudflare Hiccups: The Generations API endpoint experienced a temporary disruption due to issues with upstream infrastructure providers, causing 404 errors for some calls.
  • The announcement indicated that the issue was related to intermittent problems with Cloudflare, but the Generations API has since been restored to a healthy state.
• Retryable Restorations: Calls to that endpoint may 404 but should be re-tryable soon.
  • The announcement assured users that the service would be restored quickly, advising them to retry any failed calls.

OpenRouter (Alex Atallah) ▷ #app-showcase (4 messages):

OpenRouter Cost Dashboard, Average Request Size, Gemini Input Token Calculation

• Cost Reports get Visualized!: A member has developed a free dashboard to visualize .csv cost reports from OpenRouter, designed to analyze data from shared accounts.
  • The dashboard, available at openroutercosts.lorenzozane.com, is planned to include additional KPIs and enhanced charts, with feedback welcome.
• Average Request Size requested in Dashboard!: A member requested the addition of average request size metrics, specifically average input tokens and average output tokens, to the OpenRouter cost dashboard.
  • The dashboard’s developer committed to adding this feature soon.
• Gemini Input Tokens trigger Weird Counts!: The developer of the dashboard noted that OpenRouter’s calculation of input tokens for Gemini’s models appears to produce unusual counts when images are included in the input.
  • They are considering seeking clarification from the OpenRouter team regarding this issue, referencing a related discussion on the Google AI Developers forum.

OpenRouter (Alex Atallah) ▷ #general (528 messages🔥🔥🔥):

Deepseek pricing, OpenRouter rate limits, Gemini banning, Using OpenRouter with RAG systems, 4.6T parameter model

• Deepseek V3.1 Public Release Imminent!: Many users eagerly await the public release of Deepseek v3.1, craving it like fent and anticipating it will be free starting in September.
• Paid Deepseek offers Faster Responses: Users confirm that paying for Deepseek models on OpenRouter results in faster response times compared to the free models, with one user switching due to Chutes slowing responses, but the user experience on the free models are not as good due to constant rate limits.
  • One user stated, ever since that thing with chutes slowing responses I just said screw it i pay for it.
• OpenRouter API Keys Vulnerable to Leaks and Exploits: A user reported a loss of $300 due to a leaked OpenRouter API key and sought advice on identifying the source of the unauthorized usage, but it’s possible for threat actors to use a proxy to mask their origin IP and the user is responsible for any leaked keys.
• Is Gemini Doing the Banning Tango?: Users report massive banning occurring on Gemini, leading many to seek alternatives and reminisce about the AI Dungeon purge caused by OpenAI.
  • One user lamented, we’re being sent back to 2023.
• OpenRouter API keys can be used in RAG?: Users discuss the possibility of using OpenRouter LLM API keys in RAG systems with locally stored vector databases created by Milvus.
  • The consensus is that it’s possible, but OpenRouter doesn’t directly support embeddings, so you’ll have to retrieve documents using milvus and put it with your prompt question to the OpenRouter LLM API.

OpenRouter (Alex Atallah) ▷ #new-models (3 messages):

“

• Readybot.io Announces OpenRouter - New Models: Readybot.io has announced updates and information regarding new models available on the OpenRouter platform.
• OpenRouter’s New Models Updates: The OpenRouter platform highlights the latest additions and changes to its selection of AI models, as announced by Readybot.io.

OpenRouter (Alex Atallah) ▷ #discussion (16 messages🔥):

Qwen3 coder 480b, DeepSeek v3 0324, Zero return from generative AI, Google Gemini 400 Error, Cohere reasoning model

• LLMs struggle to format output correctly: Users are finding that LLMs like Qwen3 coder 480b and DeepSeek v3 0324 struggle to follow instructions for formatting their output properly, often resulting in bugs and ignored prompts.
  • One user found them not useful and rather distracting, often creating tic-tac-toe sites instead of the intended application.
• Most orgs see ZERO return on Generative AI: According to an AFR Chanticleer report, 95% of organizations are getting zero return out of their generative AI deployment.
  • The report notes this is focused on companies that have deployed customized AI models, and the key problem is companies and their tech vendors are not spending enough time ensuring that their customized AI models keep learning about the nuances of their businesses.
• Google Gemini Models trigger 400 Error: Google Gemini models return HTTP 400 errors when assistant messages with tool calls use the OpenAI-standard complex content format [{"type": "text", "text": "..."}] instead of simple string format.
  • This issue affects all google/gemini-* models and only occurs when tool calls and tool results are present in the message chain.
• Cohere Releases Reasoning Model: Cohere just dropped a reasoning model with further details available on Discord.
  • No further details were available.
• Feature Request: Auto-Collapse lengthy user messages: A user requested if it’s possible to automatically collapse lengthy user messages in the chatroom.
  • The user praised the chatroom and the chat management.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Cloudflare outage, Generations API stability

• Generations API Hit by Cloudflare Hiccups: The Generations API endpoint experienced a temporary disruption due to issues with upstream infrastructure providers, causing 404 errors for some calls.
  • The announcement indicated that the issue was related to intermittent problems with Cloudflare, but the Generations API has since been restored to a healthy state.
• Retryable Restorations: Calls to that endpoint may 404 but should be re-tryable soon.
  • The announcement assured users that the service would be restored quickly, advising them to retry any failed calls.

OpenRouter (Alex Atallah) ▷ #app-showcase (4 messages):

OpenRouter Cost Dashboard, Average Request Size, Gemini Input Token Calculation

• Cost Reports get Visualized!: A member has developed a free dashboard to visualize .csv cost reports from OpenRouter, designed to analyze data from shared accounts.
  • The dashboard, available at openroutercosts.lorenzozane.com, is planned to include additional KPIs and enhanced charts, with feedback welcome.
• Average Request Size requested in Dashboard!: A member requested the addition of average request size metrics, specifically average input tokens and average output tokens, to the OpenRouter cost dashboard.
  • The dashboard’s developer committed to adding this feature soon.
• Gemini Input Tokens trigger Weird Counts!: The developer of the dashboard noted that OpenRouter’s calculation of input tokens for Gemini’s models appears to produce unusual counts when images are included in the input.
  • They are considering seeking clarification from the OpenRouter team regarding this issue, referencing a related discussion on the Google AI Developers forum.

OpenRouter (Alex Atallah) ▷ #general (528 messages🔥🔥🔥):

Deepseek pricing, OpenRouter rate limits, Gemini banning, Using OpenRouter with RAG systems, 4.6T parameter model

• Deepseek V3.1 Public Release Imminent!: Many users eagerly await the public release of Deepseek v3.1, craving it like fent and anticipating it will be free starting in September.
• Paid Deepseek offers Faster Responses: Users confirm that paying for Deepseek models on OpenRouter results in faster response times compared to the free models, with one user switching due to Chutes slowing responses, but the user experience on the free models are not as good due to constant rate limits.
  • One user stated, ever since that thing with chutes slowing responses I just said screw it i pay for it.
• OpenRouter API Keys Vulnerable to Leaks and Exploits: A user reported a loss of $300 due to a leaked OpenRouter API key and sought advice on identifying the source of the unauthorized usage, but it’s possible for threat actors to use a proxy to mask their origin IP and the user is responsible for any leaked keys.
• Is Gemini Doing the Banning Tango?: Users report massive banning occurring on Gemini, leading many to seek alternatives and reminisce about the AI Dungeon purge caused by OpenAI.
  • One user lamented, we’re being sent back to 2023.
• OpenRouter API keys can be used in RAG?: Users discuss the possibility of using OpenRouter LLM API keys in RAG systems with locally stored vector databases created by Milvus.
  • The consensus is that it’s possible, but OpenRouter doesn’t directly support embeddings, so you’ll have to retrieve documents using milvus and put it with your prompt question to the OpenRouter LLM API.

OpenRouter (Alex Atallah) ▷ #new-models (3 messages):

“

• Readybot.io Announces OpenRouter - New Models: Readybot.io has announced updates and information regarding new models available on the OpenRouter platform.
• OpenRouter’s New Models Updates: The OpenRouter platform highlights the latest additions and changes to its selection of AI models, as announced by Readybot.io.

OpenRouter (Alex Atallah) ▷ #discussion (16 messages🔥):

Qwen3 coder 480b, DeepSeek v3 0324, Zero return from generative AI, Google Gemini 400 Error, Cohere reasoning model

• LLMs struggle to format output correctly: Users are finding that LLMs like Qwen3 coder 480b and DeepSeek v3 0324 struggle to follow instructions for formatting their output properly, often resulting in bugs and ignored prompts.
  • One user found them not useful and rather distracting, often creating tic-tac-toe sites instead of the intended application.
• Most orgs see ZERO return on Generative AI: According to an AFR Chanticleer report, 95% of organizations are getting zero return out of their generative AI deployment.
  • The report notes this is focused on companies that have deployed customized AI models, and the key problem is companies and their tech vendors are not spending enough time ensuring that their customized AI models keep learning about the nuances of their businesses.
• Google Gemini Models trigger 400 Error: Google Gemini models return HTTP 400 errors when assistant messages with tool calls use the OpenAI-standard complex content format [{"type": "text", "text": "..."}] instead of simple string format.
  • This issue affects all google/gemini-* models and only occurs when tool calls and tool results are present in the message chain.
• Cohere Releases Reasoning Model: Cohere just dropped a reasoning model with further details available on Discord.
  • No further details were available.
• Feature Request: Auto-Collapse lengthy user messages: A user requested if it’s possible to automatically collapse lengthy user messages in the chatroom.
  • The user praised the chatroom and the chat management.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Cloudflare outage, Generations API stability

• Generations API Hit by Cloudflare Hiccups: The Generations API endpoint experienced a temporary disruption due to issues with upstream infrastructure providers, causing 404 errors for some calls.
  • The announcement indicated that the issue was related to intermittent problems with Cloudflare, but the Generations API has since been restored to a healthy state.
• Retryable Restorations: Calls to that endpoint may 404 but should be re-tryable soon.
  • The announcement assured users that the service would be restored quickly, advising them to retry any failed calls.

OpenRouter (Alex Atallah) ▷ #app-showcase (4 messages):

OpenRouter Cost Dashboard, Average Request Size, Gemini Input Token Calculation

• Cost Reports get Visualized!: A member has developed a free dashboard to visualize .csv cost reports from OpenRouter, designed to analyze data from shared accounts.
  • The dashboard, available at openroutercosts.lorenzozane.com, is planned to include additional KPIs and enhanced charts, with feedback welcome.
• Average Request Size requested in Dashboard!: A member requested the addition of average request size metrics, specifically average input tokens and average output tokens, to the OpenRouter cost dashboard.
  • The dashboard’s developer committed to adding this feature soon.
• Gemini Input Tokens trigger Weird Counts!: The developer of the dashboard noted that OpenRouter’s calculation of input tokens for Gemini’s models appears to produce unusual counts when images are included in the input.
  • They are considering seeking clarification from the OpenRouter team regarding this issue, referencing a related discussion on the Google AI Developers forum.

OpenRouter (Alex Atallah) ▷ #general (528 messages🔥🔥🔥):

Deepseek pricing, OpenRouter rate limits, Gemini banning, Using OpenRouter with RAG systems, 4.6T parameter model

• Deepseek V3.1 Public Release Imminent!: Many users eagerly await the public release of Deepseek v3.1, craving it like fent and anticipating it will be free starting in September.
• Paid Deepseek offers Faster Responses: Users confirm that paying for Deepseek models on OpenRouter results in faster response times compared to the free models, with one user switching due to Chutes slowing responses, but the user experience on the free models are not as good due to constant rate limits.
  • One user stated, ever since that thing with chutes slowing responses I just said screw it i pay for it.
• OpenRouter API Keys Vulnerable to Leaks and Exploits: A user reported a loss of $300 due to a leaked OpenRouter API key and sought advice on identifying the source of the unauthorized usage, but it’s possible for threat actors to use a proxy to mask their origin IP and the user is responsible for any leaked keys.
• Is Gemini Doing the Banning Tango?: Users report massive banning occurring on Gemini, leading many to seek alternatives and reminisce about the AI Dungeon purge caused by OpenAI.
  • One user lamented, we’re being sent back to 2023.
• OpenRouter API keys can be used in RAG?: Users discuss the possibility of using OpenRouter LLM API keys in RAG systems with locally stored vector databases created by Milvus.
  • The consensus is that it’s possible, but OpenRouter doesn’t directly support embeddings, so you’ll have to retrieve documents using milvus and put it with your prompt question to the OpenRouter LLM API.

OpenRouter (Alex Atallah) ▷ #new-models (3 messages):

“

• Readybot.io Announces OpenRouter - New Models: Readybot.io has announced updates and information regarding new models available on the OpenRouter platform.
• OpenRouter’s New Models Updates: The OpenRouter platform highlights the latest additions and changes to its selection of AI models, as announced by Readybot.io.

OpenRouter (Alex Atallah) ▷ #discussion (16 messages🔥):

Qwen3 coder 480b, DeepSeek v3 0324, Zero return from generative AI, Google Gemini 400 Error, Cohere reasoning model

• LLMs struggle to format output correctly: Users are finding that LLMs like Qwen3 coder 480b and DeepSeek v3 0324 struggle to follow instructions for formatting their output properly, often resulting in bugs and ignored prompts.
  • One user found them not useful and rather distracting, often creating tic-tac-toe sites instead of the intended application.
• Most orgs see ZERO return on Generative AI: According to an AFR Chanticleer report, 95% of organizations are getting zero return out of their generative AI deployment.
  • The report notes this is focused on companies that have deployed customized AI models, and the key problem is companies and their tech vendors are not spending enough time ensuring that their customized AI models keep learning about the nuances of their businesses.
• Google Gemini Models trigger 400 Error: Google Gemini models return HTTP 400 errors when assistant messages with tool calls use the OpenAI-standard complex content format [{"type": "text", "text": "..."}] instead of simple string format.
  • This issue affects all google/gemini-* models and only occurs when tool calls and tool results are present in the message chain.
• Cohere Releases Reasoning Model: Cohere just dropped a reasoning model with further details available on Discord.
  • No further details were available.
• Feature Request: Auto-Collapse lengthy user messages: A user requested if it’s possible to automatically collapse lengthy user messages in the chatroom.
  • The user praised the chatroom and the chat management.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Cloudflare outage, Generations API stability

• Generations API Hit by Cloudflare Hiccups: The Generations API endpoint experienced a temporary disruption due to issues with upstream infrastructure providers, causing 404 errors for some calls.
  • The announcement indicated that the issue was related to intermittent problems with Cloudflare, but the Generations API has since been restored to a healthy state.
• Retryable Restorations: Calls to that endpoint may 404 but should be re-tryable soon.
  • The announcement assured users that the service would be restored quickly, advising them to retry any failed calls.

OpenRouter (Alex Atallah) ▷ #app-showcase (4 messages):

OpenRouter Cost Dashboard, Average Request Size, Gemini Input Token Calculation

• Cost Reports get Visualized!: A member has developed a free dashboard to visualize .csv cost reports from OpenRouter, designed to analyze data from shared accounts.
  • The dashboard, available at openroutercosts.lorenzozane.com, is planned to include additional KPIs and enhanced charts, with feedback welcome.
• Average Request Size requested in Dashboard!: A member requested the addition of average request size metrics, specifically average input tokens and average output tokens, to the OpenRouter cost dashboard.
  • The dashboard’s developer committed to adding this feature soon.
• Gemini Input Tokens trigger Weird Counts!: The developer of the dashboard noted that OpenRouter’s calculation of input tokens for Gemini’s models appears to produce unusual counts when images are included in the input.
  • They are considering seeking clarification from the OpenRouter team regarding this issue, referencing a related discussion on the Google AI Developers forum.

OpenRouter (Alex Atallah) ▷ #general (528 messages🔥🔥🔥):

Deepseek pricing, OpenRouter rate limits, Gemini banning, Using OpenRouter with RAG systems, 4.6T parameter model

• Deepseek V3.1 Public Release Imminent!: Many users eagerly await the public release of Deepseek v3.1, craving it like fent and anticipating it will be free starting in September.
• Paid Deepseek offers Faster Responses: Users confirm that paying for Deepseek models on OpenRouter results in faster response times compared to the free models, with one user switching due to Chutes slowing responses, but the user experience on the free models are not as good due to constant rate limits.
  • One user stated, ever since that thing with chutes slowing responses I just said screw it i pay for it.
• OpenRouter API Keys Vulnerable to Leaks and Exploits: A user reported a loss of $300 due to a leaked OpenRouter API key and sought advice on identifying the source of the unauthorized usage, but it’s possible for threat actors to use a proxy to mask their origin IP and the user is responsible for any leaked keys.
• Is Gemini Doing the Banning Tango?: Users report massive banning occurring on Gemini, leading many to seek alternatives and reminisce about the AI Dungeon purge caused by OpenAI.
  • One user lamented, we’re being sent back to 2023.
• OpenRouter API keys can be used in RAG?: Users discuss the possibility of using OpenRouter LLM API keys in RAG systems with locally stored vector databases created by Milvus.
  • The consensus is that it’s possible, but OpenRouter doesn’t directly support embeddings, so you’ll have to retrieve documents using milvus and put it with your prompt question to the OpenRouter LLM API.

OpenRouter (Alex Atallah) ▷ #new-models (3 messages):

“

• Readybot.io Announces OpenRouter - New Models: Readybot.io has announced updates and information regarding new models available on the OpenRouter platform.
• OpenRouter’s New Models Updates: The OpenRouter platform highlights the latest additions and changes to its selection of AI models, as announced by Readybot.io.

OpenRouter (Alex Atallah) ▷ #discussion (16 messages🔥):

Qwen3 coder 480b, DeepSeek v3 0324, Zero return from generative AI, Google Gemini 400 Error, Cohere reasoning model

• LLMs struggle to format output correctly: Users are finding that LLMs like Qwen3 coder 480b and DeepSeek v3 0324 struggle to follow instructions for formatting their output properly, often resulting in bugs and ignored prompts.
  • One user found them not useful and rather distracting, often creating tic-tac-toe sites instead of the intended application.
• Most orgs see ZERO return on Generative AI: According to an AFR Chanticleer report, 95% of organizations are getting zero return out of their generative AI deployment.
  • The report notes this is focused on companies that have deployed customized AI models, and the key problem is companies and their tech vendors are not spending enough time ensuring that their customized AI models keep learning about the nuances of their businesses.
• Google Gemini Models trigger 400 Error: Google Gemini models return HTTP 400 errors when assistant messages with tool calls use the OpenAI-standard complex content format [{"type": "text", "text": "..."}] instead of simple string format.
  • This issue affects all google/gemini-* models and only occurs when tool calls and tool results are present in the message chain.
• Cohere Releases Reasoning Model: Cohere just dropped a reasoning model with further details available on Discord.
  • No further details were available.
• Feature Request: Auto-Collapse lengthy user messages: A user requested if it’s possible to automatically collapse lengthy user messages in the chatroom.
  • The user praised the chatroom and the chat management.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Activity Analytics API, Allowed Models API, OpenRouter Developer APIs

• OpenRouter releases Activity Analytics API: OpenRouter announced the release of a new Activity Analytics API that allows users to programmatically retrieve their daily activity rollups, as well as those for their organizations, documented here.
• OpenRouter lists allowed Models API: OpenRouter has released an Allowed Models API, enabling users and organizations to programmatically fetch the models they are permitted to access based on their provider restrictions, documented here.

OpenRouter (Alex Atallah) ▷ #general (211 messages🔥🔥):

OpenWebUI Memory Feature, GPT-5 Context Issues, Deepseek v3.1 Availability on OpenRouter, Stealth Model Speculation (Grok-4 Code), Free Model Options on OpenRouter

• Decoding OpenWebUI’s Memory Lane: In OpenWebUI, the memory feature defaults to manual input, but addons exist to automatically save relevant conversation snippets, enhancing recall.
  • Relevant memories are injected into the system prompt, influencing the model’s responses by providing contextually appropriate information.
• GPT-5’s Context Crisis: Token Tumbles: Users report issues with openai/gpt-5’s 400k context window, with calls failing at ~66k tokens, yielding a silent 200 OK response and 0 tokens.
  • When trying gpt-5-chat with Cline on Cursor, users also experienced an exceeded context window error at <100k tokens.
• Deepseek v3.1 Delay: OpenRouter’s Holdout: Users are questioning why Deepseek v3.1 is not yet available on OpenRouter, even as other providers like Chutes offer it, though the latter has the base version.
  • OpenRouter is reportedly awaiting an official announcement before launching Deepseek v3.1, suggesting a provider-dependent release strategy.
• Stealth Model Speculation: Grok-4 Code Emerges: Speculation points to Grok-4 Code as the stealth model used by Cline and Cursor, with one member giving it a 90% chance of being accurate.
  • A member hinted that Deepseek instruct was another possible contender.
• Unearthing Free Model Trove on OpenRouter: OpenRouter offers a range of free models, including Llama 3.3-70B-Instruct:free, accessible via the Together AI models page.
  • Users are advised to navigate the Together AI models page and search for models labeled as “free”.

OpenRouter (Alex Atallah) ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter (Alex Atallah) ▷ #discussion (12 messages🔥):

LLMs Formatting Output, AFR Chanticleer AI Report, Google Gemini Models, OpenAI-standard complex content format, tool calling flows

• LLMs Struggle with Output Formatting: Users reported having bad experiences using LLMs like Qwen3 coder 480b and DeepSeek v3 0324 due to their inability to follow formatting instructions properly.
  • The outputs often had bugs, failed to display, and frequently ignored the initial prompt, instead creating unrelated content like a tic tac toe site.
• AI Report Causes Market Nervousness: An AI report suggests that 95% of organizations are getting zero return from their generative AI deployments, particularly those using customized AI models.
  • The report indicates that companies aren’t spending enough time ensuring their customized AI models keep learning, while a shadow AI economy has developed where employees rely on general AI models like ChatGPT and Gemini.
• Google Gemini Models Return 400 Error: Google Gemini models return an HTTP 400 error when assistant messages with tool calls use the OpenAI-standard complex content format [{"type": "text", "text": "..."}] instead of a simple string format.
  • This issue affects all google/gemini-* models but does not impact openai/* or anthropic/* models, and only occurs when tool calls and tool results are present in the message chain.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Chutes Capacity, Server Outage

• Chutes Capacity Plummets Offline: The Chutes Capacity service experienced an outage, with their servers going offline.
  • The team is actively working on restoring the servers and anticipates commencing recovery efforts shortly.
• Quick Recovery Anticipated for Chutes Capacity: Engineers are on standby to initiate the recovery process for Chutes Capacity as soon as the servers are back online.
  • No estimated time of full service restoration was given.

OpenRouter (Alex Atallah) ▷ #general (638 messages🔥🔥🔥):

DeepSeek outage, Chutes overload, OpenRouter pricing, Alternatives to DeepSeek, BYOK 5% fee

• DeepSeek v3 Outage Angers Users: Users report DeepSeek v3 is experiencing frequent internal server errors and rate limits, with some unable to generate outputs even after multiple attempts, with one user saying it’s so slow that it’s genuinely just not generating anything but I’m not getting any error messages.
  • Some speculate that Chutes, a primary provider for DeepSeek on OpenRouter, is experiencing issues due to high demand, leading to provider errors and slow performance.
• Chutes Overload Blamed for DeepSeek Issues: Several members are reporting that the overload causes 429 errors, suggesting that Chutes is experiencing a bottleneck due to miners not ramping up to meet demand; one member noted that it was completely fine all day until like 30 min ago.
  • There’s speculation that Chutes may be intentionally rate-limiting the OpenRouter API key to encourage users to purchase credits directly from them, with one user advising to just burn your credits and never use their service again.
• OpenRouter Pricing Debated Amidst Outages: With DeepSeek models barely working, some users are questioning the value of paying for OpenRouter, particularly as they are still getting rate-limited, with users expressing that a 10 USD investment for 1k free messages/day for a free model is no longer a good deal.
  • One user suggested that users with only one model in mind should’ve gone directly for the models directly, such as with DeepSeek, which may have automatic caching on their API, and further stating that the 10 USD would have lasted for months anyway.
• Free Model Alternatives Sought: Users are recommending alternative free models such as Dolphin 3.0 Mistral 24B and Mistral nemo; the latter being described as super similar to DeepSeek.
  • Some users also mentioned Z.AI: GLM 4.5 Air (free), for work related stuff, but needing a prompt; finally one user hopes for Qwen3 235B A22B (free) hosted somewhere.
• OpenRouter BYOK comes with 5% Fee: Members discovered that OpenRouter charges a 5% fee even when users bring their own API key (BYOK), leading to a discussion about whether this is a fair practice.
  • One user joked Greedy /jor 5% when you bring your own key, with another member responding you’re welcome not to use it lol.

OpenRouter (Alex Atallah) ▷ #discussion (35 messages🔥):

OpenRouter File API Integration, Tool Calling Accuracy Stats, Qwen3 32B Pricing, DeepInfra Turbo Endpoint, New Providers Section UI

• OpenRouter should Integrate File API: A member suggested that OpenRouter should figure out how to integrate a files API, noting that the top 3 labs already have this feature.
  • No further discussion was made.
• Tool Calling Accuracy: More Control Needed: A member shared thoughts on tool calling accuracy stats, arguing that the setup and environment needs to be more controlled for accurate comparison with confidence intervals.
  • They added that the apps, tools, and use cases can be vastly different, making it pointless to compare the tool call success rate without more rigor.
• Qwen3 32B priced absurdly low: Members noticed low pricing for Qwen3 32B on Chutes at $0.018/$0.072 MTok in/out, same with Mistral Small.
  • It was noted that the 32b dense version is cheaper than the moe 30b a3 version, prompting some disappointment about the lack of good providers for 30A3B.
• DeepInfra Throughput Claim Discrepancy: A member noted DeepInfra on Maverick does 600+ TPS (fp8) but another one said OR says DeepInfra runs at 83 TPS with a maximum of 105 TPS.
  • The second member clarified that they were referring to the DeepInfra Turbo endpoint.
• Providers Section Sparks UI Feedback: A member questioned if the new Providers section was bothering anyone else, mentioning that everything blurs together with the spacing, font size and separation feeling wrong.
  • Another member agreed that it looks a bit weird, but thinks it is just because it’s new and unfamiliar.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Chutes Capacity, Server Outage

• Chutes Capacity Plummets Offline: The Chutes Capacity service experienced an outage, with their servers going offline.
  • The team is actively working on restoring the servers and anticipates commencing recovery efforts shortly.
• Quick Recovery Anticipated for Chutes Capacity: Engineers are on standby to initiate the recovery process for Chutes Capacity as soon as the servers are back online.
  • No estimated time of full service restoration was given.

OpenRouter (Alex Atallah) ▷ #general (638 messages🔥🔥🔥):

DeepSeek outage, Chutes overload, OpenRouter pricing, Alternatives to DeepSeek, BYOK 5% fee

• DeepSeek v3 Outage Angers Users: Users report DeepSeek v3 is experiencing frequent internal server errors and rate limits, with some unable to generate outputs even after multiple attempts, with one user saying it’s so slow that it’s genuinely just not generating anything but I’m not getting any error messages.
  • Some speculate that Chutes, a primary provider for DeepSeek on OpenRouter, is experiencing issues due to high demand, leading to provider errors and slow performance.
• Chutes Overload Blamed for DeepSeek Issues: Several members are reporting that the overload causes 429 errors, suggesting that Chutes is experiencing a bottleneck due to miners not ramping up to meet demand; one member noted that it was completely fine all day until like 30 min ago.
  • There’s speculation that Chutes may be intentionally rate-limiting the OpenRouter API key to encourage users to purchase credits directly from them, with one user advising to just burn your credits and never use their service again.
• OpenRouter Pricing Debated Amidst Outages: With DeepSeek models barely working, some users are questioning the value of paying for OpenRouter, particularly as they are still getting rate-limited, with users expressing that a 10 USD investment for 1k free messages/day for a free model is no longer a good deal.
  • One user suggested that users with only one model in mind should’ve gone directly for the models directly, such as with DeepSeek, which may have automatic caching on their API, and further stating that the 10 USD would have lasted for months anyway.
• Free Model Alternatives Sought: Users are recommending alternative free models such as Dolphin 3.0 Mistral 24B and Mistral nemo; the latter being described as super similar to DeepSeek.
  • Some users also mentioned Z.AI: GLM 4.5 Air (free), for work related stuff, but needing a prompt; finally one user hopes for Qwen3 235B A22B (free) hosted somewhere.
• OpenRouter BYOK comes with 5% Fee: Members discovered that OpenRouter charges a 5% fee even when users bring their own API key (BYOK), leading to a discussion about whether this is a fair practice.
  • One user joked Greedy /jor 5% when you bring your own key, with another member responding you’re welcome not to use it lol.

OpenRouter (Alex Atallah) ▷ #discussion (35 messages🔥):

OpenRouter File API Integration, Tool Calling Accuracy Stats, Qwen3 32B Pricing, DeepInfra Turbo Endpoint, New Providers Section UI

• OpenRouter should Integrate File API: A member suggested that OpenRouter should figure out how to integrate a files API, noting that the top 3 labs already have this feature.
  • No further discussion was made.
• Tool Calling Accuracy: More Control Needed: A member shared thoughts on tool calling accuracy stats, arguing that the setup and environment needs to be more controlled for accurate comparison with confidence intervals.
  • They added that the apps, tools, and use cases can be vastly different, making it pointless to compare the tool call success rate without more rigor.
• Qwen3 32B priced absurdly low: Members noticed low pricing for Qwen3 32B on Chutes at $0.018/$0.072 MTok in/out, same with Mistral Small.
  • It was noted that the 32b dense version is cheaper than the moe 30b a3 version, prompting some disappointment about the lack of good providers for 30A3B.
• DeepInfra Throughput Claim Discrepancy: A member noted DeepInfra on Maverick does 600+ TPS (fp8) but another one said OR says DeepInfra runs at 83 TPS with a maximum of 105 TPS.
  • The second member clarified that they were referring to the DeepInfra Turbo endpoint.
• Providers Section Sparks UI Feedback: A member questioned if the new Providers section was bothering anyone else, mentioning that everything blurs together with the spacing, font size and separation feeling wrong.
  • Another member agreed that it looks a bit weird, but thinks it is just because it’s new and unfamiliar.

OpenRouter (Alex Atallah) ▷ #announcements (1 messages):

Chutes Capacity, Server Outage

• Chutes Capacity Plummets Offline: The Chutes Capacity service experienced an outage, with their servers going offline.
  • The team is actively working on restoring the servers and anticipates commencing recovery efforts shortly.
• Quick Recovery Anticipated for Chutes Capacity: Engineers are on standby to initiate the recovery process for Chutes Capacity as soon as the servers are back online.
  • No estimated time of full service restoration was given.

OpenRouter (Alex Atallah) ▷ #general (638 messages🔥🔥🔥):

DeepSeek outage, Chutes overload, OpenRouter pricing, Alternatives to DeepSeek, BYOK 5% fee

• DeepSeek v3 Outage Angers Users: Users report DeepSeek v3 is experiencing frequent internal server errors and rate limits, with some unable to generate outputs even after multiple attempts, with one user saying it’s so slow that it’s genuinely just not generating anything but I’m not getting any error messages.
  • Some speculate that Chutes, a primary provider for DeepSeek on OpenRouter, is experiencing issues due to high demand, leading to provider errors and slow performance.
• Chutes Overload Blamed for DeepSeek Issues: Several members are reporting that the overload causes 429 errors, suggesting that Chutes is experiencing a bottleneck due to miners not ramping up to meet demand; one member noted that it was completely fine all day until like 30 min ago.
  • There’s speculation that Chutes may be intentionally rate-limiting the OpenRouter API key to encourage users to purchase credits directly from them, with one user advising to just burn your credits and never use their service again.
• OpenRouter Pricing Debated Amidst Outages: With DeepSeek models barely working, some users are questioning the value of paying for OpenRouter, particularly as they are still getting rate-limited, with users expressing that a 10 USD investment for 1k free messages/day for a free model is no longer a good deal.
  • One user suggested that users with only one model in mind should’ve gone directly for the models directly, such as with DeepSeek, which may have automatic caching on their API, and further stating that the 10 USD would have lasted for months anyway.
• Free Model Alternatives Sought: Users are recommending alternative free models such as Dolphin 3.0 Mistral 24B and Mistral nemo; the latter being described as super similar to DeepSeek.
  • Some users also mentioned Z.AI: GLM 4.5 Air (free), for work related stuff, but needing a prompt; finally one user hopes for Qwen3 235B A22B (free) hosted somewhere.
• OpenRouter BYOK comes with 5% Fee: Members discovered that OpenRouter charges a 5% fee even when users bring their own API key (BYOK), leading to a discussion about whether this is a fair practice.
  • One user joked Greedy /jor 5% when you bring your own key, with another member responding you’re welcome not to use it lol.

OpenRouter (Alex Atallah) ▷ #discussion (35 messages🔥):

OpenRouter File API Integration, Tool Calling Accuracy Stats, Qwen3 32B Pricing, DeepInfra Turbo Endpoint, New Providers Section UI

• OpenRouter should Integrate File API: A member suggested that OpenRouter should figure out how to integrate a files API, noting that the top 3 labs already have this feature.
  • No further discussion was made.
• Tool Calling Accuracy: More Control Needed: A member shared thoughts on tool calling accuracy stats, arguing that the setup and environment needs to be more controlled for accurate comparison with confidence intervals.
  • They added that the apps, tools, and use cases can be vastly different, making it pointless to compare the tool call success rate without more rigor.
• Qwen3 32B priced absurdly low: Members noticed low pricing for Qwen3 32B on Chutes at $0.018/$0.072 MTok in/out, same with Mistral Small.
  • It was noted that the 32b dense version is cheaper than the moe 30b a3 version, prompting some disappointment about the lack of good providers for 30A3B.
• DeepInfra Throughput Claim Discrepancy: A member noted DeepInfra on Maverick does 600+ TPS (fp8) but another one said OR says DeepInfra runs at 83 TPS with a maximum of 105 TPS.
  • The second member clarified that they were referring to the DeepInfra Turbo endpoint.
• Providers Section Sparks UI Feedback: A member questioned if the new Providers section was bothering anyone else, mentioning that everything blurs together with the spacing, font size and separation feeling wrong.
  • Another member agreed that it looks a bit weird, but thinks it is just because it’s new and unfamiliar.

OpenRouter (Alex Atallah) ▷ #announcements (2 messages):

Self-Serve Refunds, Activity Improvements, Token Usage Breakdown, 3rd party credit usage, Chutes Capacity Offline

• Refunds are now self-serve!: Users can now instantly refund accidental credit purchases made within 24 hours for non-crypto purchases, according to this post.
  • This feature aims to provide more immediate control over billing errors.
• Activity Page gets Supercharged: The activity page now displays token usage broken down by token type and includes 3rd party credit usage, as announced here.
  • These improvements offer greater visibility into usage patterns and costs.
• Chutes Capacity Offline: Chutes Capacity went offline, but the team is actively working on bringing their servers back online, with recovery expected to start soon.
  • Users were informed of the issue and the ongoing efforts to restore service.

OpenRouter (Alex Atallah) ▷ #app-showcase (1 messages):

Deno or-models tool, OpenRouter model list

• Deno Tool Gets Bug Fixes: A member has fixed bugs and cleaned the output of their Deno or-models tool, which is used to inspect the OpenRouter model list.
  • The tool features a local 24-hour cache to prevent spamming the API; the tool is available here.
• Deno Update Required: To get the latest version of the or-models tool, users need to run a specific command.
  • The command is deno run -r -A jsr:@fry69/or-models --version because Deno does not automatically update to the latest version.

OpenRouter (Alex Atallah) ▷ #general (525 messages🔥🔥🔥):

Deepseek v3 issues and outages, Chutes' rate limiting and API key management, Alternatives to Deepseek for roleplaying, Azure credits liquidation strategies, Sonnet 4 pricing inconsistencies

• Deepseek V3 Server Gets Deep Fried: Users are reporting widespread issues with Deepseek V3, including internal server errors and rate limiting, particularly affecting those using it for roleplaying on platforms like Janitor AI.
  • Many attribute the issues to Chutes, a primary provider for Deepseek, struggling to meet demand and implementing stricter rate limits due to over-gooning.
• Chutes Blamed for Deepseek Outages: Many users suspect Chutes of intentionally rate-limiting OpenRouter’s API key to encourage users to purchase credits directly from them, leading to frustration and calls to boycott OpenRouter.
  • While paid requests reportedly still work, the situation has sparked debate about the ethics of Chutes’ actions and OpenRouter’s silence on the matter, with some suggesting OpenRouter should find an alternative provider.
• Roleplaying AI Models: Deepseek, Mistral, and Llama: When Deepseek is down, users recommend exploring alternative models for roleplaying, such as Mistral and Llama, with some mentioning a free Dolphin3.0 Mistral 24B model as a workable option.
  • Others suggest trying Deepseek R1, but there are conflicting reports on whether it’s also experiencing similar issues.
• Azure Credit Cash Conversion Conundrum: A user is seeking advice on converting approximately $40,000 USD in Azure credits to cash following their startup’s shutdown, acknowledging the potential issues with selling credits due to liability concerns.
  • Suggestions range from selling the credits as AI inference credits to a lighthearted offer of $50 for the entire amount, with a warning about potential risks from the buyer’s actions.
• Sonnet 4’s Erratic Pricing Escalates Concerns: Users are reporting inconsistent pricing with the Sonnet 4 endpoint on OpenRouter, experiencing sudden spikes in costs (10x) for calls using the same amount of tokens.
  • The community is requesting separate endpoints for Sonnet 4 and the 1M token version to avoid unexpected cost increases.

OpenRouter (Alex Atallah) ▷ #discussion (16 messages🔥):

Self-serve refunds, Chatroom app creation, Qwen Coder via Cerebras, Tool call evals, Model Select performance

• Refunds are now self-service!: Users can now self-serve refunds of unused credits within 24 hours of purchase; an announcement is coming soon.
• Chatroom app creation on OpenRouter!: Members are discussing new chatroom app creation features on OpenRouter.
• Qwen Coder + Cerebras = 🔥: The combination of Qwen Coder and Cerebras is gaining attention, particularly for coding related tasks.
• Tool call evals in progress!: OpenRouter is actively working on tool call evals (link to tweet).
• Model Selection = Faster!: The model selection process has been improved, resulting in a faster experience to open and better search functionality.

OpenRouter (Alex Atallah) ▷ #app-showcase (3 messages):

The Last RAG (TLRAG), NoChain Orchestrator, Statelessness & Digital Amnesia, Persistent Identity, Token Costs

• Last RAG Solves Stateless LLMs: The user is building The Last RAG (TLRAG), a foundational LLM architecture that solves problems like statelessness and digital amnesia, lacking a genuine persistent identity, massive token costs, the context window war, and expensive fine-tuning cycles.
  • TLRAG introduces a persistent, long-term memory system combined with a Dynamic Work Space (DWS), to allow the AI to curate its history and remember key interactions, decisions, and emotional contexts.
• TLRAG Gives AI Persistent Identity: TLRAG gives the AI a persistent identity core—the “Heart”, which is a living document shaped by its own synthesized experiences and memories, allowing it to develop a consistent, authentic, and self-aware personality over time.
  • It also leverages a “Window Flush” mechanism to assemble a lean, intelligent dossier with only the most relevant short-term and long-term memories, yielding validated cost savings of up to 98% over long conversations compared to standard RAG.
• NoChain Orchestrator replaces Frameworks: The NoChain Orchestrator takes the core concepts of TLRAG and makes them robust and reliable for production, replacing complex, unpredictable agent frameworks with a deterministic, server-side control plane.
  • It uses hard-coded logic to manage memory, context, and tool use, eliminating the “black box” nature of many agentic systems and delivering predictable, reliable, and testable AI behavior, and more info can be found in their blogpost.
• Explore TLRAG and NoChain concepts: The user shared a few links to explore the concepts behind TLRAG and NoChain, including a blog post and a visual pitch deck.
  • This highlights the shift from stateless tools to stateful, persistent AI partners that learn and evolve.

OpenRouter (Alex Atallah) ▷ #general (262 messages🔥🔥):

Sonnet update, tool use and structured output with open source models, GPT-5 performance, Gemini 3 as a disappointment, OpenRouter Image resizing

• OSS Models Struggle with Tool Use: Members discussed the challenge that many high-end open source models don’t support tool use, structured output, or response format, which are critical for many applications.
  • It was noted that while some providers might not support it, the model itself often does, and prompts can be used to enable tool usage, though with potential accuracy tradeoffs.
• Debate on GPT-5’s Hallucinations and Performance: Discussion arose around GPT-5, with some praising it as the first model to push SOTA forward while dropping hallucinations and improving alignment, suggesting it’s a step towards AGI.
  • Others were more critical, with one member claiming that GPT-5 is the worst and that GPT-4.1 mini was better.
• Image Resizing on OpenRouter: Only in Chatroom: A member asked if OpenRouter resizes images on the fly before sending them to the LLM.
  • It was clarified that image resizing only happens in the chatroom, and otherwise, the image is passed through without modification.
• Navigating GPT-OSS-120B on Cerebras via OpenRouter: A user shared a comprehensive guide on effectively using gpt-oss-120b on Cerebras through OpenRouter, emphasizing that guiding the output through the prompt is key to achieving consistent, schema-clean JSON.
  • The guide includes a working configuration, a Python implementation example, and notes on what doesn’t work, such as using the /completions endpoint or setting response_format.
• Copilot vs Cline/Kilo with OpenRouter: Members discussed different tools such as Copilot, Cline, Kilo, and Roo to use OpenRouter API.
  • It was discussed that Cline/Kilo are better with OR, also that Copilot has options to use OpenRouter, and in their chat tab it’s supposedly less likely to edit code and talks more, but haven’t used it

OpenRouter (Alex Atallah) ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter (Alex Atallah) ▷ #discussion (8 messages🔥):

GPU Rentals, AI TLD issues, Chatroom Caching

• GPU Rental Rave: Runpod, Prime Intellect, Modal: A member suggests to rent some GPUs from Runpod, Prime Intellect and Modal to experiment before investing in Macs.
  • They linked to a post on X: ArtificialAnlys.
• AI TLDs Cause API Endpoint Concern: Members expressed worry about AI companies changing their API endpoint due to AI TLD issues.
  • It’s not exactly a generic TLD.
• Cache Chat: Explicit Caching Proposed: A member suggests adding a cache button in the chatroom to explicitly cache a message.
  • The goal is to let users explicitly cache messages.

OpenRouter (Alex Atallah) ▷ #general (800 messages🔥🔥🔥):

GPT-5 vs GPT-5 Chat, Gemini 3.0 vs GPT-5, Deepseek Switching to Ascend, Horizon Beta Replacement

• GPT-5 Reasoning Debate Erupts: Users debate the difference between GPT-5 and GPT-5 Chat, with some suggesting GPT-5 Chat has less reasoning capabilities and is safer, while others point out that GPT-5 requires a key and GPT-5-chat does not.
  • Some suggest using gpt-5-explainer to explain the differences to friends and family, while others find GPT-5 chat to have ZERO reasoning capabilities.
• Google Poised to Pounce with Genie 3: Members express that Google is poised to win the AI race, considering it created the transformer and has the infrastructure, budget, and talent to succeed, with Genie 3 touted as crazy cool.
  • Some members look forward to Gemini 3.0 wiping the floor with GPT-5, while others point out that Google’s .0 models are not that good.
• Deepseek R2 on Ascend is Approaching: A user reported that Deepseek is switching to Ascend and launching R2, which might provide a performance boost for the model.
  • Some members express hope that Deepseek will be way better, while others share that past Deepseek models were just too unhinged.
• Horizon Beta Replaced by GPT-5 Family: The AI model Horizon Beta has been replaced by GPT-5, with no option to revert, causing disappointment among users who found it useful.
  • Some speculate that Horizon was early versions of GPT-5, and that free users will be directed to GPT-5 after they run out of free requests.

OpenRouter (Alex Atallah) ▷ #new-models (2 messages):

“

• No significant activity: The channel shows no significant discussion or new model announcements.
  • No topics warrant summarization based on the provided message history.
• Channel Inactivity: The provided message history for the OpenRouter - New Models channel appears to be empty.
  • There are no discussions, links, or announcements to summarize at this time.

OpenRouter (Alex Atallah) ▷ #discussion (23 messages🔥):

GPT-5 BYOK, o3, OpenRouter Trusted Partner, generation_time, moderation_latency

• GPT-5 to be BYOK?: A member asked if GPT-5 will always be BYOK-only like o3 on OpenRouter.
• OpenRouter’s role as trusted partner: A member congratulated OpenRouter on being one of OpenAI’s most trusted partners for the new series release.
  • They mentioned how much of an impact GPT-4 has had on the world and how much Gemini 2.5 has had in the dev sphere, and how cool OR has been to watch as a product.
• generation_time’s inclusion of other latencies: A member asked if generation_time includes moderation_latency and/or latency.
  • They also asked if latency includes moderation_latency and noted that the OpenRouter API documentation is vague on this.
• Gemini has PDF reading issues: Members reported that Gemini is not able to read the PDF files via URL while Sonnet can, even with examples from the OpenRouter multimodal documentation.
• Files API troubles: A member expressed the need for OR to figure out Files API, citing that switching between providers when you want to use Files API is a pain.

OpenRouter (Alex Atallah) ▷ #general (800 messages🔥🔥🔥):

GPT-5 vs GPT-5 Chat, Gemini 3.0 vs GPT-5, Deepseek Switching to Ascend, Horizon Beta Replacement

• GPT-5 Reasoning Debate Erupts: Users debate the difference between GPT-5 and GPT-5 Chat, with some suggesting GPT-5 Chat has less reasoning capabilities and is safer, while others point out that GPT-5 requires a key and GPT-5-chat does not.
  • Some suggest using gpt-5-explainer to explain the differences to friends and family, while others find GPT-5 chat to have ZERO reasoning capabilities.
• Google Poised to Pounce with Genie 3: Members express that Google is poised to win the AI race, considering it created the transformer and has the infrastructure, budget, and talent to succeed, with Genie 3 touted as crazy cool.
  • Some members look forward to Gemini 3.0 wiping the floor with GPT-5, while others point out that Google’s .0 models are not that good.
• Deepseek R2 on Ascend is Approaching: A user reported that Deepseek is switching to Ascend and launching R2, which might provide a performance boost for the model.
  • Some members express hope that Deepseek will be way better, while others share that past Deepseek models were just too unhinged.
• Horizon Beta Replaced by GPT-5 Family: The AI model Horizon Beta has been replaced by GPT-5, with no option to revert, causing disappointment among users who found it useful.
  • Some speculate that Horizon was early versions of GPT-5, and that free users will be directed to GPT-5 after they run out of free requests.

OpenRouter (Alex Atallah) ▷ #new-models (2 messages):

“

• No significant activity: The channel shows no significant discussion or new model announcements.
  • No topics warrant summarization based on the provided message history.
• Channel Inactivity: The provided message history for the OpenRouter - New Models channel appears to be empty.
  • There are no discussions, links, or announcements to summarize at this time.

OpenRouter (Alex Atallah) ▷ #discussion (23 messages🔥):

GPT-5 BYOK, o3, OpenRouter Trusted Partner, generation_time, moderation_latency

• GPT-5 to be BYOK?: A member asked if GPT-5 will always be BYOK-only like o3 on OpenRouter.
• OpenRouter’s role as trusted partner: A member congratulated OpenRouter on being one of OpenAI’s most trusted partners for the new series release.
  • They mentioned how much of an impact GPT-4 has had on the world and how much Gemini 2.5 has had in the dev sphere, and how cool OR has been to watch as a product.
• generation_time’s inclusion of other latencies: A member asked if generation_time includes moderation_latency and/or latency.
  • They also asked if latency includes moderation_latency and noted that the OpenRouter API documentation is vague on this.
• Gemini has PDF reading issues: Members reported that Gemini is not able to read the PDF files via URL while Sonnet can, even with examples from the OpenRouter multimodal documentation.
• Files API troubles: A member expressed the need for OR to figure out Files API, citing that switching between providers when you want to use Files API is a pain.

OpenRouter (Alex Atallah) ▷ #general (254 messages🔥🔥):

GPT-OSS performance woes, Quantization Levels, Qwen3 Coder Removal, DeepSeek structured output

• GPT-OSS Model Bashed for Poor Performance: Members in the channel derided the GPT-OSS models, citing that even smaller models are better, with one stating that the 120B model is “dead on arrival” after someone had an initial experience resulting in a “really ugly typo in the headline”.
  • One member linked to a Reddit thread summing up the sentiment that it’s a “dud model” and more like a publicity stunt than a useful model.
• Provider Routing Allows Custom Quantization Levels: When a user asked how to avoid quantized models, one member pointed out that users can configure quantization levels using the provider routing feature, noting models are excluded if the provider doesn’t meet the quantization level.
  • The user then recommended using FP8 to avoid the quantized models, noting that anything under that is “worse than useless”.
• Qwen3-Coder:Free Tier Gets the Axe: Multiple users noted that the Qwen3-Coder:Free has been removed and is no longer available through any providers.
  • Members lamented the loss and mentioned that they hoped it would return.
• DeepSeek’s JSON output support is provider-dependent: Users discussed the inconsistent support for structured output (JSON) with DeepSeek-r1, noting that while it’s supported on their own API, it may vary on OpenRouter depending on the provider.
  • One member linked to a Reddit thread and a filtered view of OpenRouter models that support structured outputs, with most agreeing that this is provider specific.
• SDK Downgrade Fixes Reasoning Issue: A user experienced that reasoning tokens were being duplicated when using the GPT-OSS model, which was resolved by downgrading the SDK from version 0.7.3 to 0.6.0.
  • A team member confirmed that the fix is in the main branch, and linked to the pull request stating it has not been cut into a release yet, and that they will release the fix soon.

OpenRouter (Alex Atallah) ▷ #discussion (29 messages🔥):

20 Questions Benchmark, GPT-OSS Hallucinations, OpenRouter Provider Sanity Checks, Harmony Format and Identity, Tool Use Validation

• 20 Questions Benchmark hits Kaggle: A member developed a 20 Questions benchmark and found a similar competition on Kaggle, although the Kaggle competition was for custom agents.
  • Their 2.5 Pro agent achieved 8/20 words on their benchmark.
• GPT-OSS Under Fire For Hallucinations: GPT-OSS is reported to be prone to hallucination, making it a potentially unsuitable choice for certain applications.
  • A member suggested that GPT-4.1 is a much safer choice, especially with prompt/context engineering.
• OpenRouter Mulls Provider Sanity Checks: There is a suggestion for OpenRouter to implement sanity checks or smoke tests for all providers, focusing on formatting and tool call evaluation.
  • Providers failing the test could be temporarily removed from the serving pool, and there is acknowledgement that current checks are relatively simple but more thorough solutions are in progress.
• Harmony Format and Identity Face Scrutiny: A member inquired about how system and developer messages are treated via the OpenRouter API, specifically whether they are interpreted as developer messages or model_identity for gpt-oss.
  • They linked to a Discord message regarding the topic of harmony format, identity vs system / developer message (discord.com).
• Tool Use Validation Under Development: Automatic validation of tool-use, distinguishing between good and bad implementations, is under development as a better solution.
  • This is related to a tweet (x.com) discussing the same topic.

OpenRouter (Alex Atallah) ▷ #general (254 messages🔥🔥):

GPT-OSS performance woes, Quantization Levels, Qwen3 Coder Removal, DeepSeek structured output

• GPT-OSS Model Bashed for Poor Performance: Members in the channel derided the GPT-OSS models, citing that even smaller models are better, with one stating that the 120B model is “dead on arrival” after someone had an initial experience resulting in a “really ugly typo in the headline”.
  • One member linked to a Reddit thread summing up the sentiment that it’s a “dud model” and more like a publicity stunt than a useful model.
• Provider Routing Allows Custom Quantization Levels: When a user asked how to avoid quantized models, one member pointed out that users can configure quantization levels using the provider routing feature, noting models are excluded if the provider doesn’t meet the quantization level.
  • The user then recommended using FP8 to avoid the quantized models, noting that anything under that is “worse than useless”.
• Qwen3-Coder:Free Tier Gets the Axe: Multiple users noted that the Qwen3-Coder:Free has been removed and is no longer available through any providers.
  • Members lamented the loss and mentioned that they hoped it would return.
• DeepSeek’s JSON output support is provider-dependent: Users discussed the inconsistent support for structured output (JSON) with DeepSeek-r1, noting that while it’s supported on their own API, it may vary on OpenRouter depending on the provider.
  • One member linked to a Reddit thread and a filtered view of OpenRouter models that support structured outputs, with most agreeing that this is provider specific.
• SDK Downgrade Fixes Reasoning Issue: A user experienced that reasoning tokens were being duplicated when using the GPT-OSS model, which was resolved by downgrading the SDK from version 0.7.3 to 0.6.0.
  • A team member confirmed that the fix is in the main branch, and linked to the pull request stating it has not been cut into a release yet, and that they will release the fix soon.

OpenRouter (Alex Atallah) ▷ #discussion (29 messages🔥):

20 Questions Benchmark, GPT-OSS Hallucinations, OpenRouter Provider Sanity Checks, Harmony Format and Identity, Tool Use Validation

• 20 Questions Benchmark hits Kaggle: A member developed a 20 Questions benchmark and found a similar competition on Kaggle, although the Kaggle competition was for custom agents.
  • Their 2.5 Pro agent achieved 8/20 words on their benchmark.
• GPT-OSS Under Fire For Hallucinations: GPT-OSS is reported to be prone to hallucination, making it a potentially unsuitable choice for certain applications.
  • A member suggested that GPT-4.1 is a much safer choice, especially with prompt/context engineering.
• OpenRouter Mulls Provider Sanity Checks: There is a suggestion for OpenRouter to implement sanity checks or smoke tests for all providers, focusing on formatting and tool call evaluation.
  • Providers failing the test could be temporarily removed from the serving pool, and there is acknowledgement that current checks are relatively simple but more thorough solutions are in progress.
• Harmony Format and Identity Face Scrutiny: A member inquired about how system and developer messages are treated via the OpenRouter API, specifically whether they are interpreted as developer messages or model_identity for gpt-oss.
  • They linked to a Discord message regarding the topic of harmony format, identity vs system / developer message (discord.com).
• Tool Use Validation Under Development: Automatic validation of tool-use, distinguishing between good and bad implementations, is under development as a better solution.
  • This is related to a tweet (x.com) discussing the same topic.

OpenRouter (Alex Atallah) ▷ #announcements (5 messages):

Anthropic Opus 4.1, OpenAI returns to Open Source, GPT-OSS models

• New Anthropic Opus 4.1 takes the Crown: The latest model, Anthropic Opus 4.1, is now live and tops the charts in SWE Bench, the leading coding benchmark, as announced on X.
  • It can be accessed here.
• OpenAI’s GPT-OSS Models Debut: OpenAI is returning to open source with the launch of gpt-oss, new open-weight models with variable reasoning, with OpenRouter as a launch partner, as announced on X.
  • Two models are available: gpt-oss-120b at $0.15/M input tokens and $0.60/M output tokens, and gpt-oss-20b at $0.05/M input tokens and $0.20/M output tokens.
• GPT-OSS-20B price correction: A user pointed out that the prices were mislabeled.
  • The listing was updated to reflect the prices on the 20b model.

OpenRouter (Alex Atallah) ▷ #app-showcase (1 messages):

gardasio: ChatGPT.com https://x.com/Gardasio/status/1952501913586442541

OpenRouter (Alex Atallah) ▷ #general (251 messages🔥🔥):

Model vs Models Prioritization, Gemini video understanding, Claude providers caching, Qwen-image model, GPTs agents training

• Model vs Models Prioritization Issue Fixed: A member reported that the issue with model vs models prioritization has been fixed via a quick fix that only uses model or models but not both, as per this link.
• OpenRouter Not Stripping Cache Params?: Some members reported that Claude providers don’t support caching, and they expect OpenRouter to strip the cache params automatically.
  • One member specified that *“a few Claude providers don’t support caching, azure and google, I’ve blacklisted them in settings, but I expected openrouter to just strip the cache params automatically.”
• Qwen-Image Model incoming?: A member inquired about the possibility of getting the Qwen-image model featured on OpenRouter, pointing to the Qwen Image blogpost.
• Free Gemini V3 model questions: Members clarified that with a $10 investment, OpenRouter grants 1000 free messages a day for the free V3 model.
  • Others warned that although OpenRouter may allow 1000 requests a day, Chutes may not and that retrying requests is always an option.
• Fal.ai for Image Generation?: A member asked about a good ‘Openrouter for images’ single endpoint for image gen? and other members recommended Fal.ai and local generation using ComfyUI.
  • Fal.ai was praised for having a full API and well-structured data specs, and that a single 3060 is enough for local generation.

OpenRouter (Alex Atallah) ▷ #new-models (3 messages):

“

• No new models news to report: There were no substantial discussion points or links shared regarding new models in the OpenRouter Discord channel.
• Silence on the New Models Front: The ‘new-models’ channel appears to be quiet, with no specific updates, discussions, or links shared during the specified timeframe.

OpenRouter (Alex Atallah) ▷ #discussion (30 messages🔥):

LLM Emotional Understanding Benchmarks, EQ Benchmark, Gemma 3 27b, OCR Engine Comparisons, Sonnet Self-Moderation

• Gemma 3 Shows Emotional Intelligence: Members discussed benchmarks for LLMs in understanding human emotions, with one suggesting the EQ benchmark by EQbench.
  • It was also suggested that Gemma 3 27b is a good model for understanding emotions, while cautioning against using DeepSeek R1 for that purpose.
• Users React to Sonnet Self-Moderation: Users discuss the recent changes in Sonnet self-moderation, and whether a self-moderated option will remain available.
  • One user questioned whether LlamaGuard will still be applied to Anthropic endpoints, suggesting that it’s an industry-leading approach to protecting users, while another said they could no longer trigger it but haven’t tried very hard.
• OCR Engine Preferences Debate: A user expressed dissatisfaction with a particular OCR engine and stated they find them shit and prefer OLMo.
  • Another user asked why and another user explained that OLMo is cheaper and better on almost all benchmarks.
• DeepInfra Price Hike Angers Users: A user expressed frustration with DeepInfra for raising prices.
  • Another user confirmed that DeepInfra will cost about $1/1000 tokens, equivalent to the Mistral API, while OpenRouter is $2/1000, suggesting OpenRouter could adjust its pricing.
• GPT-OSS-120B Ready to Launch: A provider has deployed gpt-oss-120b with 65K context ready to go on launch.
  • It was noted that the moderation is applied on Anthropic and Bedrock providers, while vertex is unfiltered as per an attached image.

OpenRouter (Alex Atallah) ▷ #app-showcase (11 messages🔥):

PyrenzAI launch, Personality.gg, OpenRouter PKCE, PyrenzAI feedback

• Personality.gg enables Roleplay via OpenRouter: Personality.gg launched a roleplay site using OpenRouter for most models, providing access to all 400 models through OpenRouter PKCE (Proof Key for Code Exchange) completely free/cheap.
• PyrenzAI Launches a Free AI Chat Website: A developer announced the launch of PyrenzAI, an AI chat website with a clean UI, models, a memory system and free RAG (Retrieval-Augmented Generation) for all tiers, using OpenRouter as the main AI generation backend.
• PyrenzAI app faces speed and security critiques: A user critiqued the newly launched PyrenzAI app, noting it’s cooked in terms of both speed and security, with laggy performance and excessive fetching of user preferences (over 200+ times on every load).
• UI and UX lauded on PyrenzAI release: A member complimented the UI/UX of PyrenzAI, appreciating its unique look and style, and distinctive sidebar design compared to other apps.

OpenRouter (Alex Atallah) ▷ #general (242 messages🔥🔥):

API Errors, Deepseek r1, Free Models, Horizon Alpha, API Key credit limit

• API Errors plague OpenRouter Users: Some users reported experiencing API errors when trying to use models via the OpenRouter API, including no endpoint found errors and other issues.
  • A member suggested checking the model ID prefix and the base URL for potential misconfiguration.
• Deepseek v3 Outage Strikes Users: Users reported issues with the Deepseek v3 0324 free model, including internal errors, empty responses, and timeouts.
  • One member noted that switching to the paid version of the model resolved the issues, suggesting the free version was overloaded: free is completely overloaded. paid has none of these issue, and the actual content quality is better.
• Free Model Limits Frustrate OpenRouter Users: Several users inquired about free models with higher message limits, with one user asking if there was any free model that wont stop at 50 messages?
  • Members clarified that topping up with $10 provides a 1000 requests/day limit and referenced OpenRouter documentation detailing the limits.
• Horizon Alpha Raves Gain Momentum: Users discussed the Horizon Alpha model, with some reporting that it was reasoning effectively and offering good performance.
  • The model itself reported that it was developed by OpenAI, though other members clarified that it was likely a distilled model.
• Budget Overruns Baffle API users: A user reported being charged significantly over their API key credit limit, suspecting that running API calls in parallel with Python threads might be the cause.
  • Other users shared similar experiences, suggesting that the credit limit updates might not be real-time, leading to occasional overcharges.

OpenRouter (Alex Atallah) ▷ #new-models (1 messages):

Readybot.io: OpenRouter - New Models

OpenRouter (Alex Atallah) ▷ #discussion (23 messages🔥):

Groq OpenBench, Provider Benchmarks, GPQA Evals, Inspect.ai, Prompt Caching for Kimi K2 and GLM 4.5

• OpenBench Groqs for Provider Benchmarks: Members discussed the Groq OpenBench repository and how many times it has been posted regarding provider benchmarks.
  • One member mentioned they are already working on evals (recently got prioritized), such as GPQA per provider, and expanding to other things.
• Inspect.ai Discovery Praised: A member expressed happiness in discovering inspect.ai through the OpenBench link, noting it’s exactly what I’ve been looking for.
  • This same user noted concerns about the chat UI using their full name from their account without control over it, leading to potential doxxing.
• Prompt Caching Questioned for Kimi K2 and GLM 4.5: A user inquired whether OpenRouter supports prompt caching for Kimi K2 and GLM 4.5, noting that Moonshot’s platform directly supports it.
  • They stated it somewhat looks like it on z.ai.
• Bypassing 20MB Limit: Bigger PDFs are now sendable: Members questioned whether new feature would bypass the 20MB limit, and they mentioned that they recently added a way to send bigger pdfs.
  • The new limit is the upstream provider limit.

OpenRouter (Alex Atallah) ▷ #app-showcase (11 messages🔥):

PyrenzAI launch, Personality.gg, OpenRouter PKCE, PyrenzAI feedback

• Personality.gg enables Roleplay via OpenRouter: Personality.gg launched a roleplay site using OpenRouter for most models, providing access to all 400 models through OpenRouter PKCE (Proof Key for Code Exchange) completely free/cheap.
• PyrenzAI Launches a Free AI Chat Website: A developer announced the launch of PyrenzAI, an AI chat website with a clean UI, models, a memory system and free RAG (Retrieval-Augmented Generation) for all tiers, using OpenRouter as the main AI generation backend.
• PyrenzAI app faces speed and security critiques: A user critiqued the newly launched PyrenzAI app, noting it’s cooked in terms of both speed and security, with laggy performance and excessive fetching of user preferences (over 200+ times on every load).
• UI and UX lauded on PyrenzAI release: A member complimented the UI/UX of PyrenzAI, appreciating its unique look and style, and distinctive sidebar design compared to other apps.

OpenRouter (Alex Atallah) ▷ #general (242 messages🔥🔥):

API Errors, Deepseek r1, Free Models, Horizon Alpha, API Key credit limit

• API Errors plague OpenRouter Users: Some users reported experiencing API errors when trying to use models via the OpenRouter API, including no endpoint found errors and other issues.
  • A member suggested checking the model ID prefix and the base URL for potential misconfiguration.
• Deepseek v3 Outage Strikes Users: Users reported issues with the Deepseek v3 0324 free model, including internal errors, empty responses, and timeouts.
  • One member noted that switching to the paid version of the model resolved the issues, suggesting the free version was overloaded: free is completely overloaded. paid has none of these issue, and the actual content quality is better.
• Free Model Limits Frustrate OpenRouter Users: Several users inquired about free models with higher message limits, with one user asking if there was any free model that wont stop at 50 messages?
  • Members clarified that topping up with $10 provides a 1000 requests/day limit and referenced OpenRouter documentation detailing the limits.
• Horizon Alpha Raves Gain Momentum: Users discussed the Horizon Alpha model, with some reporting that it was reasoning effectively and offering good performance.
  • The model itself reported that it was developed by OpenAI, though other members clarified that it was likely a distilled model.
• Budget Overruns Baffle API users: A user reported being charged significantly over their API key credit limit, suspecting that running API calls in parallel with Python threads might be the cause.
  • Other users shared similar experiences, suggesting that the credit limit updates might not be real-time, leading to occasional overcharges.