Mistral introduces Voxtral 3 transcription and speech models along with a voice agent API

French AI company Mistral announced the Voxtral 3 family of audio models and a new API for voice agents on October 20. Per its changelog, four models are in public preview. Voxtral Mini Transcribe 3 and Voxtral Small Transcribe 3 handle batch transcription, Voxtral Mini Transcribe Realtime 3 handles live transcription, and Voxtral TTS 3 generates speech. Transcription covers more than 50 languages, including all 24 official EU languages. The 4-billion-parameter Mini costs $0.0035 per minute, Small $0.004 and the realtime model $0.008. Speech output covers 35+ languages and 50+ dialects at $13 per million characters.

Mistral says Small outperforms Whisper, ElevenLabs’ Scribe v2 and Gemini 3.5 Transcribe. On the company’s diarization benchmark (cpWER, lower is better), Small scores 18.4%, against 30.9% for the previous Voxtral Mini Transcribe 2. The realtime model runs at less than 500 milliseconds of latency. Mistral says its TTS costs less to run than comparable models from OpenAI, Deepgram, ElevenLabs Flash and Gemini Flash. A Pro Voice feature reproduces a speaker’s style across languages from 30 minutes of audio.

In the voice agent SDK, built on LiveKit, the realtime model listens, a model such as Mistral Small reasons and calls tools, and the TTS model replies. Mistral says each turn completes in under one second and the agent stops talking when the caller cuts in. The agent API costs $0.04 per minute with Mistral Small 4 and $0.08 with a larger model. Agents connect over SIP for phone calls and WebRTC for browsers and apps. The models are served from Mistral’s European, US and global endpoints and can also be self-deployed.