Mistral Voxtral 3: Voice Agents for $0.04 a Minute

Mistral Voxtral 3 is the French lab's bid to own the whole voice agent, not just the transcript. Launched on October 20, it bundles new speech-to-text models, a more expressive text-to-speech model and a voice agent SDK built on LiveKit. Mistral says each conversational turn completes in under one second, and everything can run on its EU-hosted API or on your own servers.

What actually happened

Mistral's announcement lays out a four-part problem: a good voice agent must listen accurately, reason about intent, speak naturally and take actions, all at once and in real time. Voxtral 3 is its answer, a family of models plus a pipeline that glues them together.

On the listening side, Voxtral Small Transcribe 3 handles multilingual transcription, diarization (who said what) and context biasing, where you feed in key terms such as product names to improve recall. Mistral says it outperforms Whisper, ElevenLabs' Scribe v2 and Gemini 3.5 Transcribe on its benchmarks. The Mini variant has 4B parameters. A realtime model targets sub-500ms latency with end-of-turn detection, so agents know when you have stopped talking.

On the speaking side, Voxtral Mini TTS 3 adjusts emotion, pace and emphasis based on the conversation. Voice presets ship in 18 languages, and a Pro Voice feature can adapt to a specific speaker from about 30 minutes of audio.

How the Mistral Voxtral 3 agent works

The SDK runs speech in, reasoning and speech out as one streaming pipeline. The realtime transcriber listens, a general model such as Mistral Small reasons and calls tools, and the speech model replies in the tone of the call so far. Mistral says no stage waits for the one before it, which is how it gets turns under a second, and the agent stops speaking when the caller cuts in.

During a call the agent can verify identity, check a bill or book an appointment through function calling, web search and retrieval. It can answer and place phone calls over SIP and run in browsers and apps over WebRTC. Because each stage is a separate model, you can swap the LLM, the voice or the language without rebuilding the agent.

The pitch in one line: a full phone agent for four cents a minute, hosted in Europe.

The fine print on benchmarks

All the "state-of-the-art" claims come from Mistral's own charts, several of which use an AI model as the judge, and the post notes which ones are internal evals. Mistral also says its TTS costs less to run than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. That is a pricing claim, not an independent quality test.

There are small inconsistencies too: the announcement says the realtime model covers more than 35 languages in one place and 40+ in another. Expect the docs to settle on one number. For the broader TTS race, see our open TTS leaderboard piece and the HeyGen voice TTS launch.

Why it matters

Voice is where AI agents meet real customers: contact centers, service desks, booking lines. Data rules make this hard in Europe, because recorded calls are sensitive. Mistral leaning on "sovereign deployment" and self-hosting is aimed squarely at banks, insurers and public bodies that cannot ship audio to a U.S. cloud.

It also squeezes pure-play voice startups. When a model lab sells transcription, reasoning and TTS in one bill, standalone providers have to win on quality or price. Big Tech is pushing in-house voice too, as with Microsoft's MAI voice models.

What to do as a reader

If you run a support line or build voice apps, test Mistral Voxtral 3 against your current stack on your own recordings, especially accented speech, numbers and names, which Mistral itself highlights as hard cases. Measure end-to-end turn time on real phone lines, not just model latency.

There is no token or Polymarket market here, so nothing to trade. Treat the benchmark charts as a reason to test, not as proof.

Sources: Mistral Voxtral 3 announcement and Mistral audio docs.