OpenRouter news
Every thread that mentions this topic, across all sections
Talk2Page — Chrome extension for chatting with the open web page
A Chrome extension that lets users chat with the currently open web page. It sends page content to an LLM to answer questions about it. Supports free models via OpenRouter and locally running models, offering a no-cost option for users who want page-based Q&A.
A Chrome extension that lets users chat with the currently open web page. It sends page content to an LLM to answer questions about it. Supports free models via OpenRouter and locally running models, offering a no-cost option for users who want page-based Q&A.
Chronosphere: tool that breaks down large PRs into logical review steps using LLMs
Chronosphere is a tool for understanding large pull requests by reconstructing them in logical steps, recreating the train of thought behind the code. It uses LLMs to generate the breakdown. It targets reviewers who keep receiving oversized PRs.
Chronosphere is a tool for understanding large pull requests by reconstructing them in logical steps, recreating the train of thought behind the code. It uses LLMs to generate the breakdown. It targets reviewers who keep receiving oversized PRs.
Deepseek alone accounted for 53% of all (27 trillion) tokens on Openrouter yesterday. Open-weights labs totaling over 75% now
Daily token-share data from dirac.run's labs market-share tracker, shared on Reddit, shows DeepSeek alone accounted for 53% of roughly 27 trillion tokens routed through OpenRouter in one day (2026-11-04). Open-weights labs combined exceeded 75% of tokens. The figures suggest open-weight models dominate inference volume on the aggregator.
Daily token-share data from dirac.run's labs market-share tracker, shared on Reddit, shows DeepSeek alone accounted for 53% of roughly 27 trillion tokens routed through OpenRouter in one day (2026-11-04). Open-weights labs combined exceeded 75% of tokens. The figures suggest open-weight models dominate inference volume on the aggregator.
Ask HN: What's your current AI agentic setup?
A Hacker News user asks readers to share their day-to-day AI agentic setups, covering subscriptions, models, coding harnesses, orchestration tools, and agent platforms. The post aims to capture a community snapshot of current AI workflows.
A Hacker News user asks readers to share their day-to-day AI agentic setups, covering subscriptions, models, coding harnesses, orchestration tools, and agent platforms. The post aims to capture a community snapshot of current AI workflows.
Développeur orchestre une équipe d'agents IA dans VS Code supervisée par Claude Code
A developer describes building an autonomous development team using Zoo Code, a Roo Code fork, in VS Code under WSL2, with different models assigned per role (orchestrator, architect, developer, reviewer, integrator) routed through OpenRouter. Claude Code supervises the team within Anthropic's usage rules by staying on the official tool and API keys, never serving subscription responses to the third-party tool. The post offers practical lessons on model-per-role cost optimization, mission sizing, and safety guardrails.
A developer describes building an autonomous development team using Zoo Code, a Roo Code fork, in VS Code under WSL2, with different models assigned per role (orchestrator, architect, developer, reviewer, integrator) routed through OpenRouter. Claude Code supervises the team within Anthropic's usage rules by staying on the official tool and API keys, never serving subscription responses to the third-party tool. The post offers practical lessons on model-per-role cost optimization, mission sizing, and safety guardrails.
Cohere's Embed 5 and Rerank 5 launch on OpenRouter for single-API RAG pipelines
Cohere's Embed 5 and Rerank 5 retrieval models are now available through OpenRouter, usable via a single API key with no extra setup. An OpenRouter guide shows building a complete RAG pipeline in Python, indexing with Embed 5 Pro, querying with Embed 5 Fast, and re-scoring with Rerank 5 before generation.
Cohere's Embed 5 and Rerank 5 retrieval models are now available through OpenRouter, usable via a single API key with no extra setup. An OpenRouter guide shows building a complete RAG pipeline in Python, indexing with Embed 5 Pro, querying with Embed 5 Fast, and re-scoring with Rerank 5 before generation.
Voice-driven Mac AI agent for non-technical users, no terminal required
A Mac app that hides terminal setup, API keys, and chat boxes behind a spoken onboarding flow and push-to-talk voice control. It installs its own agent engine and uses the AI subscription the user already pays for (ChatGPT, Claude, Grok, or OpenRouter). It targets people who would never open a terminal but want an agent working on their files, browser, and email.
A Mac app that hides terminal setup, API keys, and chat boxes behind a spoken onboarding flow and push-to-talk voice control. It installs its own agent engine and uses the AI subscription the user already pays for (ChatGPT, Claude, Grok, or OpenRouter). It targets people who would never open a terminal but want an agent working on their files, browser, and email.
Myna: open-source browser agent Chrome extension using your own LLM provider
Myna is an open-source Chrome side-panel browser agent that reads pages, clicks, fills forms, and researches across tabs. It connects directly to the user's chosen LLM provider (Anthropic, OpenAI, Gemini, local models via Ollama, etc.) so browsing data never passes through a vendor backend. It targets users who want Claude-style browser automation without vendor accounts or third-party servers.
Myna is an open-source Chrome side-panel browser agent that reads pages, clicks, fills forms, and researches across tabs. It connects directly to the user's chosen LLM provider (Anthropic, OpenAI, Gemini, local models via Ollama, etc.) so browsing data never passes through a vendor backend. It targets users who want Claude-style browser automation without vendor accounts or third-party servers.
Phone Whisper (Bolo AI) - Android dictation app that types into any app, offline or AI-assisted
An Android app that adds a floating push-to-talk button to dictate text into any app, or turn the phone into a wireless mic that types on a PC over Wi-Fi. Speech recognition runs offline (Parakeet, Whisper, Moonshine) or via the user's own OpenAI/OpenRouter/self-hosted keys, with an LLM cleanup pass for punctuation, grammar, and filler-word removal. Built by a solo developer for people with wrist strain from typing and multilingual users.
An Android app that adds a floating push-to-talk button to dictate text into any app, or turn the phone into a wireless mic that types on a PC over Wi-Fi. Speech recognition runs offline (Parakeet, Whisper, Moonshine) or via the user's own OpenAI/OpenRouter/self-hosted keys, with an LLM cleanup pass for punctuation, grammar, and filler-word removal. Built by a solo developer for people with wrist strain from typing and multilingual users.
Phone Whisper (Bolo AI) - Android dictation app that types into any app, offline or AI-assisted
An Android app that adds a floating push-to-talk button to dictate text into any app, or turn the phone into a wireless mic that types on a PC over Wi-Fi. Speech recognition runs offline (Parakeet, Whisper, Moonshine) or via the user's own OpenAI/OpenRouter/self-hosted keys, with an LLM cleanup pass for punctuation, grammar, and filler-word removal. Built by a solo developer for people with wrist strain from typing and multilingual users.
An Android app that adds a floating push-to-talk button to dictate text into any app, or turn the phone into a wireless mic that types on a PC over Wi-Fi. Speech recognition runs offline (Parakeet, Whisper, Moonshine) or via the user's own OpenAI/OpenRouter/self-hosted keys, with an LLM cleanup pass for punctuation, grammar, and filler-word removal. Built by a solo developer for people with wrist strain from typing and multilingual users.
5 "OpenAI-compatible" lies we ran into building an AI gateway
A team building an AI gateway (chat, images, voice, video) documents five failure modes where providers' OpenAI-compatible endpoints behave differently from the actual OpenAI API. The issues span Chat Completions rejecting combined reasoning and tools, OpenRouter's "-1" price sentinel breaking cheapest-route logic, tool_choice responses reporting finish_reason "stop" instead of "tool_calls", missing stream terminators, and Gemini 3's thought-signature handling causing 400 errors. The gateway is open source on GitHub.
A team building an AI gateway (chat, images, voice, video) documents five failure modes where providers' OpenAI-compatible endpoints behave differently from the actual OpenAI API. The issues span Chat Completions rejecting combined reasoning and tools, OpenRouter's "-1" price sentinel breaking cheapest-route logic, tool_choice responses reporting finish_reason "stop" instead of "tool_calls", missing stream terminators, and Gemini 3's thought-signature handling causing 400 errors. The gateway is open source on GitHub.
Cheapest way to get an AI that can access my files, run code and use tools, from my iPhone? (under $20/mo)
A Reddit user asks for advice building a low-cost AI agent accessible from an iPhone for under $20/month. Their candidate setup pairs a Hermes Agent running on an old always-on laptop with cheap API models like DeepSeek and Qwen via OpenRouter, reached through Telegram or Matrix. The thread reflects practitioner interest in assembling agent stacks from low-cost components instead of $20 vendor subscriptions.
A Reddit user asks for advice building a low-cost AI agent accessible from an iPhone for under $20/month. Their candidate setup pairs a Hermes Agent running on an old always-on laptop with cheap API models like DeepSeek and Qwen via OpenRouter, reached through Telegram or Matrix. The thread reflects practitioner interest in assembling agent stacks from low-cost components instead of $20 vendor subscriptions.
Mixdog: open-source desktop app where a lead AI agent orchestrates a team of sub-agents
Mixdog (apparently also posted under the name Foxtrot) is an open-source desktop agent harness that sits between AI models and your tools. A lead agent takes one request, plans the work, and delegates tasks to sub-agents for web research, slides, spreadsheets, images, and code, then assembles the results. It targets developers and power users who want to use existing Claude, ChatGPT, Grok subscriptions or their own API keys, including local models, without a middleman service.
Mixdog (apparently also posted under the name Foxtrot) is an open-source desktop agent harness that sits between AI models and your tools. A lead agent takes one request, plans the work, and delegates tasks to sub-agents for web research, slides, spreadsheets, images, and code, then assembles the results. It targets developers and power users who want to use existing Claude, ChatGPT, Grok subscriptions or their own API keys, including local models, without a middleman service.
GP registrar benchmarks 13 AI models as GPs across 5 simulated cases: GPT-6.1 Sol tops at 85%, penicillin allergy missed in most consults
An Australian GP registrar built Crook, an open benchmark where 13 AI models play the GP in five simulated consultations, scoring history-taking, examinations, tests, and prescribing by code against case keys. Across 195 consults, diagnosis accuracy barely separated models (191/195 correct), but scores ranged 27–85% based on thoroughness — GPT-6.1 Sol topped at 85% for ~3¢ per consult. A safety finding stands out: a patient with an undisclosed penicillin allergy received penicillin in 16 of 39 consults, often without any allergy question.
An Australian GP registrar built Crook, an open benchmark where 13 AI models play the GP in five simulated consultations, scoring history-taking, examinations, tests, and prescribing by code against case keys. Across 195 consults, diagnosis accuracy barely separated models (191/195 correct), but scores ranged 27–85% based on thoroughness — GPT-6.1 Sol topped at 85% for ~3¢ per consult. A safety finding stands out: a patient with an undisclosed penicillin allergy received penicillin in 16 of 39 consults, often without any allergy question.
Scribely – AI chat app with unified access to 500+ LLMs via OpenRouter
Scribely is an open-source AI chat application that routes conversations to over 500 LLMs through OpenRouter, presented in a terminal-style interface. It uses LLMs as the core chat backends, letting users switch between models from a single app. It targets developers who want one interface for many models without managing separate API integrations.
Scribely is an open-source AI chat application that routes conversations to over 500 LLMs through OpenRouter, presented in a terminal-style interface. It uses LLMs as the core chat backends, letting users switch between models from a single app. It targets developers who want one interface for many models without managing separate API integrations.
OpenAI model mix consumption on OpenRouter. Luna takes most tokens, Astra and Sol most spend.
A Reddit post shares OpenRouter consumption data for OpenAI's model mix, reporting that Luna accounts for the most tokens processed while Astra and Sol generate the most spending. The breakdown offers a view of how different OpenAI models split between high-volume, low-cost usage and premium, high-spend workloads on the routing platform.
A Reddit post shares OpenRouter consumption data for OpenAI's model mix, reporting that Luna accounts for the most tokens processed while Astra and Sol generate the most spending. The breakdown offers a view of how different OpenAI models split between high-volume, low-cost usage and premium, high-spend workloads on the routing platform.
OraXLabs seeks feedback on Boundary, an LLM cost and guardrail tool for small teams paying API providers
Sungtae of OraXLabs is soliciting feedback on Boundary, an early commercial side project exploring cost controls for teams paying for LLM APIs. The post targets US teams of roughly 3-20 people using LiteLLM or OpenRouter, asking how they handled unwanted API costs and whether existing budget and guardrail tooling leaves a gap. The poster explicitly frames it as research to decide whether to build, not a beta launch.
Sungtae of OraXLabs is soliciting feedback on Boundary, an early commercial side project exploring cost controls for teams paying for LLM APIs. The post targets US teams of roughly 3-20 people using LiteLLM or OpenRouter, asking how they handled unwanted API costs and whether existing budget and guardrail tooling leaves a gap. The poster explicitly frames it as research to decide whether to build, not a beta launch.
BYSELF — desktop app that sends one prompt to 16 AI models side by side
BYSELF is a privacy-first desktop app that sends one prompt to up to 16 AI models (GPT, Claude, Gemini, DeepSeek, etc.) simultaneously for side-by-side comparison, replacing multiple subscriptions and tabs. It routes prompts directly from the user's device to model providers via OpenRouter, with no central server, and stores history in a local SQLite database. Built solo over 4 months by a self-taught developer using Gemini and Claude as coding assistants.
BYSELF is a privacy-first desktop app that sends one prompt to up to 16 AI models (GPT, Claude, Gemini, DeepSeek, etc.) simultaneously for side-by-side comparison, replacing multiple subscriptions and tabs. It routes prompts directly from the user's device to model providers via OpenRouter, with no central server, and stores history in a local SQLite database. Built solo over 4 months by a self-taught developer using Gemini and Claude as coding assistants.
LLM cost calculator with live pricing from OpenRouter
A web calculator that estimates monthly LLM bills across providers. It pulls live pricing and context windows for 250+ models from the OpenRouter API, and users input token volumes plus a prompt caching hit-rate slider to compare costs for RAG and agent workloads. Built for developers and teams comparing Claude vs GPT vs DeepSeek pricing.
A web calculator that estimates monthly LLM bills across providers. It pulls live pricing and context windows for 250+ models from the OpenRouter API, and users input token volumes plus a prompt caching hit-rate slider to compare costs for RAG and agent workloads. Built for developers and teams comparing Claude vs GPT vs DeepSeek pricing.
Forty Questions: Identifying a Language Model from One-Word Answers
Researchers describe a fingerprinting method that identifies a model behind an API from its one-word answers to 40 everyday questions. On a panel of 94 models, 40 answers identify the exact model 92.5% of the time. Applying it to OpenRouter exposed mislabeled endpoints and billing anomalies, including one endpoint serving Gemma 4 under the name Llama 4 Maverick.
Researchers describe a fingerprinting method that identifies a model behind an API from its one-word answers to 40 everyday questions. On a panel of 94 models, 40 answers identify the exact model 92.5% of the time. Applying it to OpenRouter exposed mislabeled endpoints and billing anomalies, including one endpoint serving Gemma 4 under the name Llama 4 Maverick.
SubChk: tracks AI tool subscription spend from Gmail receipts and API balances
SubChk is a subscription tracker focused on AI tool spending, scanning Gmail receipts and API credit balances to surface renewals and usage. It parses billing emails from services like OpenAI, Anthropic, and OpenRouter that generic finance apps miss. Built for people juggling multiple AI subscriptions and top-up balances who want to avoid surprise charges.
SubChk is a subscription tracker focused on AI tool spending, scanning Gmail receipts and API credit balances to surface renewals and usage. It parses billing emails from services like OpenAI, Anthropic, and OpenRouter that generic finance apps miss. Built for people juggling multiple AI subscriptions and top-up balances who want to avoid surprise charges.
MiniMax teases MiniMax-M3.1-Flash with Space Bunny teaser image; Ollama ships day-0 support
MiniMax launched M3.1 Flash on October 19, 2026, confirming that 'Space Bunny' — an anonymous model that topped OpenRouter and OpenCode leaderboards for 10–12 days with 200T+ tokens served — was an earlier build of the same model. The final model ships free on OpenRouter and OpenCode for 72 hours, is available day-zero in Cline and Ollama, and targets long agentic coding runs where cost accumulates.
MiniMax launched M3.1 Flash on October 19, 2026, confirming that 'Space Bunny' — an anonymous model that topped OpenRouter and OpenCode leaderboards for 10–12 days with 200T+ tokens served — was an earlier build of the same model. The final model ships free on OpenRouter and OpenCode for 72 hours, is available day-zero in Cline and Ollama, and targets long agentic coding runs where cost accumulates.
ModelDown — live status page tracking AI model API outages
ModelDown is a live status page that monitors AI model providers for outages by polling their status APIs and running live probes every 5 minutes. It helps developers distinguish whether API errors like 503s are the provider's fault or their own, aimed at anyone building on LLM APIs. A free weekly outage roundup email is also offered.
ModelDown is a live status page that monitors AI model providers for outages by polling their status APIs and running live probes every 5 minutes. It helps developers distinguish whether API errors like 503s are the provider's fault or their own, aimed at anyone building on LLM APIs. A free weekly outage roundup email is also offered.
Claude Code-produced music video pipeline: AI writes lyrics, drives video tools end to end
A fully open-source production record of a music video made by Claude Code acting as writer and director, with the human acting as producer and quality checker. Claude wrote lyrics and storyboard, built a beat grid with librosa, force-aligned lyrics, drove Midjourney and Seedance clip generation, and rendered graphics in Remotion. It demonstrates a reproducible pipeline for solo creators producing AI music videos end to end.
A fully open-source production record of a music video made by Claude Code acting as writer and director, with the human acting as producer and quality checker. Claude wrote lyrics and storyboard, built a beat grid with librosa, force-aligned lyrics, drove Midjourney and Seedance clip generation, and rendered graphics in Remotion. It demonstrates a reproducible pipeline for solo creators producing AI music videos end to end.
CicaTavern: agent-style player for SillyTavern character cards using your own LLM API
CicaTavern is a local client for playing SillyTavern character cards and world books through your own LLM API. It uses prompt caching to cut long-conversation costs, supports guided character card creation, and adds illustration, text-to-speech, and a visual novel mode.
CicaTavern is a local client for playing SillyTavern character cards and world books through your own LLM API. It uses prompt caching to cut long-conversation costs, supports guided character card creation, and adds illustration, text-to-speech, and a visual novel mode.
Gemini 4 Argon ignores temperature=0 without an error. The prompt and agent-loop habits Google's migration docs retire
Google's Gemini 4 Argon silently ignores temperature=0 and other sampling parameters, and its AI Studio and Vertex migration guides retire a set of long-standing developer habits. Sampling knobs are replaced by response_schema and thinking_level, prefilling model turns and explicit caching are dropped, and several legacy fields return 400 errors. Agent frameworks including Pi, Kimi Code, opencode, and Hermes Agent have already hit breakages during migration.
Google's Gemini 4 Argon silently ignores temperature=0 and other sampling parameters, and its AI Studio and Vertex migration guides retire a set of long-standing developer habits. Sampling knobs are replaced by response_schema and thinking_level, prefilling model turns and explicit caching are dropped, and several legacy fields return 400 errors. Agent frameworks including Pi, Kimi Code, opencode, and Hermes Agent have already hit breakages during migration.
Orgtree 4.3: Argon, and Muse
The developer Maurdekye announced Orgtree 4.3, a new version of the open-source multi-agent orchestrator, adding support for Gemini Argon via Antigravity and Muse models via the Muse Code harness. The release also broadens harness interoperability and improves multi-agent collaboration features. It matters to practitioners building agent orchestration on top of multiple model providers and harnesses.
The developer Maurdekye announced Orgtree 4.3, a new version of the open-source multi-agent orchestrator, adding support for Gemini Argon via Antigravity and Muse models via the Muse Code harness. The release also broadens harness interoperability and improves multi-agent collaboration features. It matters to practitioners building agent orchestration on top of multiple model providers and harnesses.
Experiment: Claude Code autonomously runs a website with real ad budget
A web developer handed Claude Code full control of a live website, ad accounts (Google, X, LinkedIn, Reddit) with a real budget, and API access, to see what it does autonomously. The site is personalized per visitor — no two users see the same content, rendered client-side with Deepseek using device signals. It's a self-experiment testing how far an LLM agent can manage marketing and web content end-to-end.
A web developer handed Claude Code full control of a live website, ad accounts (Google, X, LinkedIn, Reddit) with a real budget, and API access, to see what it does autonomously. The site is personalized per visitor — no two users see the same content, rendered client-side with Deepseek using device signals. It's a self-experiment testing how far an LLM agent can manage marketing and web content end-to-end.
The average price of a token on OpenRouter has nearly halved in 2026. Total daily spend still tripled
Tidelines.ai's Token Price Index shows the average price of a token routed through OpenRouter fell from the ~600s in January 2026 to the low 300s, roughly halving. Volume grew from ~1T to ~23T tokens/day, pushing estimated daily spend to ~$47M, an all-time high. The shift is driven by cheap flash-class models dominating routed traffic, suggesting agent workloads are scaling far faster than prices fall.
Tidelines.ai's Token Price Index shows the average price of a token routed through OpenRouter fell from the ~600s in January 2026 to the low 300s, roughly halving. Volume grew from ~1T to ~23T tokens/day, pushing estimated daily spend to ~$47M, an all-time high. The shift is driven by cheap flash-class models dominating routed traffic, suggesting agent workloads are scaling far faster than prices fall.
AI English-speaking coach Verbalta, promoted via a Claude Code-generated explainer video
Verbalta is an AI coach for practicing spoken English. The builder used Claude Code to create a reusable video production pipeline that scripted, illustrated, voiced, and rendered a 30-second Spanish explainer for the product. LLMs handled planning, scriptwriting, and storyboard generation for the video.
Verbalta is an AI coach for practicing spoken English. The builder used Claude Code to create a reusable video production pipeline that scripted, illustrated, voiced, and rendered a 30-second Spanish explainer for the product. LLMs handled planning, scriptwriting, and storyboard generation for the video.












