Mistral announced Voxtral 3 on October 20, 2026, adding a speech stack for real-time voice agents that runs through its API and also supports self-deployment options such as on-prem and private cloud setups.
According to the Mistral launch post, the release includes multilingual transcription, text-to-speech, turn detection, reasoning, and tool calling in a LiveKit-based streaming pipeline, plus a voice agent SDK for connecting those pieces.
Mistral says the stack supports transcription in 50+ languages, while speech output covers 35+ languages and 50+ dialects. It also offers voice cloning from about 30 minutes of source audio. Published pricing starts at $0.0035 per transcription minute and $13 per million text-to-speech characters.
For builders, that means one vendor to test across listening, response generation, and spoken output instead of stitching together separate STT and TTS services. One caveat: Mistral’s own materials do not fully agree on realtime transcription coverage, so teams targeting specific locales should confirm language support before migrating.