OpenRouter

What this means for you: A speech-to-text model, which turns recorded audio into written text, is now available through OpenRouter's API (a way for software to call a service directly). It is Moondream's improved version of NVIDIA Parakeet TDT 0.6B v3, with the same speed and lower word error rate on every benchmark group. It handles 25 European languages and gives timestamps for segments and for individual words.

How to use it: Find it at openrouter.ai/moondream/parakeet-ultra. It costs $0.06 per hour of audio and is served by Prime Intellect.

What this means for you: OpenRouter, one account for hundreds of AI models, updated Guardrails, the controls admins use to set AI budgets, model access and content moderation. An admin can set limits for the whole company, give certain people a bigger budget or more models inside those limits, and change one policy without touching the others.

Why it matters: OpenRouter says that in a year, AI use there went from a few people per company to entire departments. Teams of 10 or more grew 70x, and keys per org tripled.

How to use it: Admins manage this in Guardrails on OpenRouter.

What this means for you: OpenRouter has made HeyGen Voice Turbo, a text-to-speech model that reads written text aloud, free until November 30. It offers over 2,900 voices in 39 languages, takes up to 5,000 characters per request, and streams 44.1 kHz audio.

How to use it: Use the free version of HeyGen Voice Turbo on OpenRouter at openrouter.ai/heygen/heygen-voice-1-turbo:free. It costs $0 and needs no credits until November 30.

What this means for you: OpenRouter added Crouton Alpha, a free reasoning model with a hidden-name (stealth) launch. It can take in text, images, audio and video, and read up to 1M tokens (small chunks of text) at once. OpenRouter describes it as built for work your agents, which are AI helpers that carry out tasks, repeat all day.

How to use it: Choose Crouton Alpha on OpenRouter. It is free to use, and OpenRouter says there is no training on your data.

What this means for you: You can turn a single image into a short video clip, 3 or 5 seconds long, at 480p or 720p. You give it a starting image and a text prompt, and it returns the clip.

How to use it: It costs from $0.005 per second of video. It is live on OpenRouter, a service that gives you one account for many AI models, at openrouter.ai/tbc/mischief-1.

What this means for you: OpenRouter is one account that lets you use many AI models. It has a new way of choosing which company runs your request. Before, it picked mostly by price. Now it weighs quality against the real cost of the request, so requests stay cheap but are also fast and reliable.

What this means for you: A new voice model that turns written text into spoken audio, built for voice agents (AI that talks with callers or customers). It offers 4,000+ voices, and the audio is served from the US or the EU.

How to use it: AssemblyAI-TTS-1 is live on OpenRouter, which gives one account and one API (a way for software to talk to a service) to many AI models. It costs $15 per 1M characters, and the audio is served from the US or EU.

What this means for you: A coding model that uses fewer tokens (the small pieces of text an AI reads and writes, which you pay for) to get the work done. Particle reports it uses about 45% fewer tokens on average than the model it was built on, DeepSeek V4.1 Flash, with accuracy close to it. So agent tasks finish sooner and cost less.

How to use it: Particle-1 is live on OpenRouter at openrouter.ai/particle-ai/particle-1. It costs $0.25 per million input tokens and $1 per million output tokens, with 1.05M context.

What this means for you: GPT-6.1 Terra from OpenAI can now be used through OpenRouter, one account that reaches many AI models. OpenRouter says it matches GPT-6 Astra on Agents' Last Exam at one-tenth the price per token, and finishes tasks over 3x faster on average than Astra and GPT-6.1 Sol. It costs $1 per million input tokens and $5 per million output tokens (tokens are small chunks of text), with a 1.05M-token context.

How to use it: Find it at openrouter.ai/openai/gpt-6.1-terra.

What this means for you: OpenRouter gives you one account to use many AI models. Its new Security Center shows all your API keys (the passwords apps use to call it) in one place, shows which are safe to remove, and sorts them by risk. You can set spend alerts, set a periodic spend limit on keys with no cap, set a maximum lifetime on every key, and control an IP allowlist.

How to use it: Open openrouter.ai/settings/security. Turn on key spend alerts, remove unused keys, set a periodic spend limit on uncapped keys, and set a max key lifetime on all keys.

What this means for you: OpenRouter now offers Flint, a coding model from a company called Recursive. Its weights are open, meaning anyone can download and run it. It is free to use at launch. Recursive built it for coding agents, which are AI helpers that write and run code on their own.

How to use it: Use the free version at openrouter.ai/recursive/flint-v1.0:free. Recursive also says the open weights are on Hugging Face and that you can try it free on opencode or pi.dev.

What this means for you: Voxtral 3 is a set of voice models from Mistral. Two turn speech into written text (transcription), and one reads text aloud. They cover more than 50 languages. Outputs are a transcript or spoken audio.

How to use it: It is available on OpenRouter, which gives one account and one API (a way for software to connect to a service) for many AI models. Pick Voxtral 3 from the model list there.

What this means for you: OpenRouter now lists MiniMax M3.1 Flash, an AI model you can use through OpenRouter's single account and API (the connection developers use to reach models). It is free for 72 hours. It is the model that ran earlier under the codename Space Bunny, which OpenRouter says was its number one model for 10 straight days and processed 56.7T tokens.

How to use it: MiniMax M3.1 Flash is on OpenRouter and free for the next 72 hours from the Oct 19 announcement.

What this means for you: A new AI model called Mercury 3 is available that writes text very fast, at more than 1,000 tokens per second (a token is a small chunk of a word). It can also take in 1M tokens at once, which is a very long stretch of text. At launch it costs $0.10 per million tokens you send in and $0.35 per million tokens it writes back, which is 50% off.

How to use it: Open OpenRouter and pick Mercury 3 from its model list. The 50% launch price applies at launch.

What this means for you: OpenRouter now lists DeepSeek V4.1 Flash 1017, an updated version of DeepSeek's V4.1 Flash model. DeepSeek says it is more token efficient, meaning it gets through work using fewer tokens. It also reports major gains on agentic benchmarks, which are tests of an AI carrying out multi-step tasks on its own.

How to use it: Open openrouter.ai/deepseek/deepseek-v4.1-flash-1017 to use it now. DeepSeek is the first provider, and more providers are on the way.

What this means for you: You can give Krea 3 one to four reference images and a written prompt, and it returns one edited image at 2K resolution. Most edits finish in under 8 seconds from start to finish.

How to use it: Krea 3 is available on OpenRouter at openrouter.ai/krea/k3-edit.

What this means for you: Span-02 is a decision model: you give it text and images, define your questions, and it returns a choice, a yes/no/unknown answer, or a score, each with probabilities. For example, it can sort items into categories you define. Its weights are open, meaning anyone can download and run it.

Why it matters: It is the second decision model OpenRouter added on Oct 16, after Perplexity's Decider models.

How to use it: Span-02 is on OpenRouter. It costs $0.01 per million input tokens (small chunks of text the model reads), output is free, and it handles 131K of context.

What this means for you: Perplexity's Decider models take text or images and return typed answers with probabilities, such as a choice or a score. Decider v1.2 is a 27B (27 billion parameter) model. Decider Flash v1.2 is a smaller 4B model that is faster and costs half as much.

How to use it: Both models are on OpenRouter at openrouter.ai/perplexity/pplx-decider-v1.2 and openrouter.ai/perplexity/pplx-decider-flash-v1.2. Open weights means anyone can download them.

What this means for you: Jev 1.14 from TypeSafe is a decision model with a better grasp of languages other than English. It can now also look at images, though that feature is experimental. Input costs $0.042 per million tokens (small chunks of text) and output is free.

How to use it: Find it on OpenRouter at openrouter.ai/typesafe/jev-1.14.

What this means for you: NVIDIA Nemotron 3.5 Super VL is an open AI model that can work with text, images, documents, app screens and video. It is made for agents, which are AI helpers that carry out long, multi-step tasks, and it can hold 1M tokens of context (a very large amount of text) at once.

How to use it: It is live on OpenRouter now.

What this means for you: Private Endpoints give a company its own reserved AI capacity and its own fine-tunes (versions of a model trained further on that company's data). Only people in that organization can see them. They are called the same way as any other model, and fall back to shared capacity when needed.

How to use it: Available today on OpenRouter Enterprise plans.

What this means for you: FLUX 3 Image Klein is an open-weights image model, meaning its files are public, and it runs on one consumer GPU (a home computer graphics card). FLUX 3 Video can now make 4K videos from a text prompt. FLUX Video Upscale can turn any video into 4K.

How to use it: All three are on OpenRouter. FLUX Video Upscale is in early access and now costs $0.28 per second, down from $0.59.

What this means for you: The ori command line tool lets you run coding assistants such as Claude Code or Codex on any model available on OpenRouter. You install it with one line typed into a terminal, then pick the model you want.

How to use it: Run curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash in a terminal. Then run a command such as ori claude --model google/gemini-4-argon. More at openrouter.ai/ori/harness.

What this means for you: You can attach one image to a chat on OpenRouter and ask the model to find things in it. Start your message with /detect {"object":"coffee cup"} to have it detect that object, or /point {"object":"croissant"} to have it point to the object.

How to use it: Moondream 3 Preview and Moondream 3.1 are now on OpenRouter, served by Prime Intellect. In chat, attach one image, then begin your message with /detect or /point followed by the object in curly braces.

What this means for you: OpenRouter's public rankings page now has a chart comparing open-weight models (models whose trained settings are published for anyone to download and run) with closed ones. This week, open-weight models handled 73% of all tokens, the small chunks of text that AI models read and write.

How to use it: Open openrouter.ai/rankings to see the chart. OpenRouter is also taking suggestions for new public charts.

What this means for you: Armor checks the actions an AI agent is about to take, called tool calls, and blocks the ones that break your rules before the agent runs them. A tool call is how an agent interacts with the outside world.

How to use it: Armor is in beta (an early version still being tested) on OpenRouter. It uses the Jev model from Typesafe AI to do the checking.

What this means for you: HeyGen Voice turns written text into spoken audio, with 10 built-in voices. HeyGen Voice Turbo is the version with a shorter delay before the audio starts.

Why it matters: HeyGen says its new voice model topped the Artificial Analysis text-to-speech leaderboard in blind tests.

How to use it: Both models are on OpenRouter through its API. API users get 50% off through Oct 31, which is $15 per million characters instead of $30, applied automatically. HeyGen Voice is also live in HeyGen itself.

What this means for you: This model turns recorded speech into written text. It makes fewer mistakes on short back-and-forth speech and handles phone numbers, addresses, emails and codes better. It also adds 14 languages, including Korean, Russian, Persian, Urdu, Zulu and Cantonese.

How to use it: It is live on OpenRouter at $0.30 per hour of audio, through AssemblyAI's Sync API (the way software sends it audio and gets text back). The model page is openrouter.ai/assemblyai/universal-3-6-pro.

What this means for you: OpenRouter, a single account that gives access to hundreds of AI models, added ScaleDown models that each do one job. Summarize turns long text, images and PDFs into short summaries. Extract takes fields you describe in plain English and returns clean JSON (a structured data format), with confidence scores and source spans. Classify-1 lets you set your own labels and returns a probability for each.

Why it matters: It comes the same day OpenRouter added several other models on Oct 13, including Liquid AI's LFM2.5-24B-A2B and Google's Gemini 4 Argon. It also follows Mercury Decide's image input on Oct 13, and Classify-1 works through OpenRouter's Decisions API.

How to use it: Find them at openrouter.ai/provider/scaledown. Launch pricing is $0.04 per million input tokens (tokens are small chunks of text) with free output, which is 20% off.

What this means for you: You can now pick Google's Gemini 4 Argon from OpenRouter, the one account that gives you many AI models. It is a Google model for long work with many steps. Google AI Studio and Google Cloud supply it.

How to use it: Go to openrouter.ai/google/gemini-4-argon to try it. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text a model reads and writes.

What this means for you: You can now send a picture along with your text to Mercury Decide and ask it typed questions about what is in the image. It answers with a choice, a score, or a yes or no with a probability attached.

Why it matters: Pricing stays where it was on Oct 8, when OpenRouter added the paid zero data retention endpoint at $0.02 per million input tokens next to the free one.

How to use it: Go to openrouter.ai/inception/mercury-decide. Put images alongside text in the state of a Decisions request. Use the free endpoint, or pay $0.02 per million input tokens with zero data retention. Images are billed as input tokens.

What this means for you: This is a reasoning model, which means it works through a problem before it answers. It is built for agents (AI that carries out tasks for you) and for using tools. Liquid AI describes it as an efficient local model for laptops and PCs, for multi-step tasks inside popular agent harnesses. It has 24 billion parameters in total but uses only 2 billion for each token it handles.

How to use it: It is now on OpenRouter. Liquid AI also published the open base and post-trained weights on Hugging Face. Open weights are model files anyone can download. Liquid AI says the model is very easy to fine-tune and customize.