Mistral AI said on 20 October that it released Voxtral 3, a family of speech models and a voice agent kit. Mistral said its realtime transcription model returns text in under 500 milliseconds in over 35 languages. It said a full agent turn (listen, reason, speak) takes under one second. The speech model costs $13 per 1M characters. Everything is available now through Mistral's API.
These figures come from Mistral's announcement. It says Whisper, Scribe v2 and Gemini 3.5 Transcribe all trailed its larger transcription model. Several voice-quality charts are labelled internal evals. Mistral also cites outside benchmarks, including Artificial Analysis, BeSimple AI's Voice Code Bench and FLEURS. The announcement shows no independent check of the Voxtral results. It gives no self-hosting price.
Mistral says no stage waits for the one before it, so the steps overlap. Its end-of-turn detection weighs sound and meaning to spot when a caller has finished, which Mistral says cuts false cut-offs. Each stage is a separate model, so a team can swap the language model or voice without rewriting the agent. That, plus a choice of where audio is processed, gives buyers more say. Before buying, ask what self-hosting costs, whether the accuracy holds on your own call audio, and what retention controls exist.