Mistral, the Paris-based open-weight lab, has released Voxtral 3, a set of transcription and voice-generation models plus an SDK for building voice agents that answer phone calls and browser sessions. Mistral AI published the announcement on 20 October, and the pricing is the sharpest part of it: a complete voice agent runs from $0.04 a minute.
The listening side is two models. Voxtral Small Transcribe 3 handles transcription across 50 languages, including the 24 official EU languages and a set of lower-resource Asian and Indic ones. It also labels who said what (diarization), accepts a list of key terms to improve recall on jargon, and predicts when a caller has stopped talking. Voxtral Mini Transcribe 3 is a 4-billion-parameter version at $0.0035 a minute for batch jobs. A separate Realtime variant targets live use with latency under 500 milliseconds in more than 35 languages, at $0.008 a minute.
The speaking side is Voxtral Mini TTS 3, which Mistral prices at $13 per million characters. It covers more than 35 languages and 50 dialects, and it picks emotion, pace, and emphasis from the surrounding conversation instead of waiting for tags. Voice presets exist in 18 languages. For a company that wants its own sound, Mistral’s Pro Voice option claims to reproduce a speaker’s character and emotional range from 30 minutes of audio, across languages.
The SDK is built on LiveKit, the real-time audio infrastructure project. It chains the realtime transcriber, a general language model such as Mistral Small, and the speech model so that no stage waits for the previous one to finish. Mistral says a full turn finishes in less than a second and that the agent stops talking the moment a person interrupts it. During a call, an agent can confirm who is phoning, look up a bill, or schedule a visit, using tool calls, web lookups, and searches of a company’s own records. They connect over SIP for phone lines and WebRTC for browsers. The three stages are independent models, so a team can swap the language model or the voice without rebuilding the agent.
Every performance claim here comes from Mistral. The company says Small Transcribe 3 beats Whisper, Scribe v2 (ElevenLabs), and Gemini 3.5 Transcribe on multilingual transcription and diarization. Its speech charts compare against a long list of rivals, but the announcement cites no independent evaluator for transcription. The text-to-speech quality charts use Voice Code Bench, a test from BeSimple AI, plus Mistral’s own evaluations, and in both cases the scoring judge is Google’s gemini-3.8-flash, a model from one of the competitors being measured. Mistral also says Voxtral TTS costs less than the rival offerings from OpenAI, Deepgram, ElevenLabs, and Google, which is a price comparison anyone can check against vendor pages.
Licensing is the open question. The announcement promises that the models are available through Mistral’s API on European, US, and global endpoints, or for self-deployment on a company’s own infrastructure, so audio never has to leave a chosen environment. It does not state a license or say whether weights can be downloaded. Full on-premises and private-cloud setups are routed to Mistral’s sales team, which suggests commercial terms rather than a free weights drop.
The $0.04 and $0.08 per-minute rates (the first with Mistral Small 4 doing the reasoning, the second with a larger model) set the number that contact-center buyers will hold against their current vendors. A company paying per call minute for a human agent or an existing voice stack now has a European, self-hostable quote to bring to the next renewal, and a reason to run a few hundred of its own recorded calls through Voxtral before trusting any chart.
Reported by Mistral AI in its Voxtral 3 announcement, published on 20 October 2026.
