Mistral released Voxtral 3 on October 20, 2026 (the date shown on Mistral’s news index), a family of speech-to-text and text-to-speech models plus a voice agent API built on LiveKit, positioned “for building agents that listen, reason, speak, and act.” It is available through Mistral’s EU-hosted API, with US and global endpoints, or for self-deployment.
The family and its list prices: Voxtral Small Transcribe 3, the most accurate transcription model, at $0.004 per minute; Voxtral Mini Transcribe 3, a 4-billion-parameter batch model, at $0.0035 per minute; Voxtral Mini Transcribe Realtime 3 at $0.008 per minute with sub-500 ms latency; and Voxtral Mini TTS 3 at $13 per million characters. The voice agent API costs $0.04 per minute with Mistral Small 4 and $0.08 per minute with a larger model. Transcription covers 50-plus languages including the 24 official EU languages, with speaker diarization, context biasing and end-of-turn detection; TTS supports voice cloning and adapts emotion, pacing and emphasis to context.
Mistral says listening, reasoning, tool calling and speaking run as one streaming pipeline, with each turn completing in under a second, over SIP or WebRTC. It reports 18.4 percent cpWER for Voxtral Small Transcribe 3 on diarization, against 30.9 percent for Scribe v2 and 38.4 percent for Gemini 3.5 Transcribe.
Why it matters: it gives European buyers a full voice-agent stack, from transcription to speech output, priced per minute and hosted in the EU.
What it does not show: all benchmark figures are Mistral’s own. The announcement does not state whether the weights are open or under what license, nor the parameter counts of the larger models, and it is inconsistent about the realtime model’s language count (35-plus in the text, 40-plus on its model card).