There are three transcription models. Voxtral Small Transcribe 3 covers 50 languages, labels who said what, and takes a list of key terms to bias toward (Mistral calls it context biasing). Voxtral Mini Transcribe 3 is a 4B model at $0.0035 a minute in batch. The realtime version costs $0.008 a minute, and Mistral says it stays under 500 ms of latency in more than 35 languages.
Voxtral Speech 3, the text-to-speech side, costs $13 per million characters and takes instructions for emotion, pacing and emphasis. Voice cloning needs 30 minutes or more of sample audio.
The voice agent SDK is built on LiveKit. The realtime transcriber listens and detects when the caller stops talking, a general model (Mistral Small, for example) reasons and calls tools, and Speech 3 answers. Mistral says each turn finishes in under a second and the agent stops talking when the caller cuts in. It handles phone calls over SIP and browser or app calls over WebRTC. The hosted agent API costs $0.04 a minute with Mistral Small 4 doing the reasoning, or $0.08 a minute with any other model.
All the benchmark charts are Mistral's own, and the TTS evals were scored by a Gemini model acting as judge. I haven't run any of it yet. It's live on Mistral's European, US and global endpoints, and self-hosting goes through their sales team.