Mistral AI introduced Voxtral 3 on Tuesday, October 20, 2026, its most complete push yet into voice. The release combines new speech recognition models, a new speech synthesis model and a software development kit meant to turn them into working voice agents for call centres, service desks and hands-free assistants. The Paris company pitches it as a sovereign option: customers can use the models through its API or run them on their own infrastructure and decide whether audio is stored.
On the listening side there are three models. Voxtral Small Transcribe 3 is the most accurate and costs $0.004 per minute. Voxtral Mini Transcribe 3, a 4-billion-parameter model, handles batch jobs for $0.0035 per minute. Both work in more than 50 languages, including all 24 official languages of the European Union, and add speaker diarization and context biasing, where users supply key terms or names to improve recognition of specialist vocabulary. A third model, Voxtral Mini Transcribe Realtime 3, is designed for live agents, returns text in under 500 milliseconds in more than 35 languages and costs $0.008 per minute.
For speech output, Voxtral Speech 3, also described as Voxtral Mini TTS 3, covers more than 35 languages and 50 dialects and adjusts emotion, pace and emphasis to the context of a conversation. It costs $13 per million characters. A Pro Voice option recreates a particular speaker's style from about 30 minutes of recorded audio, aimed at brands that want a consistent voice across languages.
The SDK is the piece that ties everything together. Built on the open-source LiveKit framework, it streams audio in, sends it to a general model such as Mistral Small for reasoning and tool calls, and speaks the answer back without waiting for each stage to finish. Mistral says a full turn takes less than a second and that the agent stops talking when a caller interrupts. Agents can check bills, verify identities or book appointments using function calls, web search and retrieval, and connect over SIP telephone lines or WebRTC in browsers and apps.
Mistral claims its Small transcription model beats OpenAI's Whisper, ElevenLabs' Scribe v2 and Google's Gemini 3.5 Transcribe on multilingual accuracy and diarization benchmarks, and that its speech model is cheaper to run than rival text-to-speech services. These comparisons come from Mistral's own evaluations, some judged automatically by another AI model, and have not yet been checked by outside testers. The Decoder's report on the launch repeats the main specifications and pricing.
The launch matters because voice is becoming a main interface for AI agents and a large commercial market, contested by OpenAI, Google, ElevenLabs, Deepgram and others. Mistral's mix of low per-minute prices, European hosting and self-deployment targets regulated industries such as healthcare, banking and the public sector, where moving customer calls to a United States cloud can be difficult.
The models are available now through Mistral's European, United States and global API endpoints, and can be tried in its Studio playgrounds. Mistral did not say whether the weights will be released openly, and running the full stack on premises requires contacting its sales team.