Mistral's Voxtral 3 is a full voice-agent stack: 50+ languages, sub-second turns, and a sovereignty pitch

• Who: Mistral AI, announced October 20, 2026 on the company’s news page.
• What: Voxtral 3 — a speech-to-text and text-to-speech model family plus a LiveKit-based voice agent SDK that runs listening, reasoning, tool calls, and speech as a single real-time streaming pipeline.
• Transcription: Voxtral Small Transcribe 3 covers 50+ languages (24 official EU languages) with speaker diarization, context biasing, and end-of-turn detection; 4.1% word error rate on FLEURS 45 per Mistral’s launch post; priced from $0.0035/min (Mini batch) to $0.008/min (Realtime).
• Speech: Voxtral Speech 3 speaks 35+ languages and 50+ dialects with emotion, pacing, and emphasis control, plus voice cloning from 30+ minutes of audio; $13 per million characters.
• Agent loop: each conversational turn completes in under one second, the agent stops talking when the caller interrupts, and it can book appointments or check bills mid-call; $0.04/min with Mistral Small 4 reasoning, $0.08/min with any other model.
• Deployment: available today through Mistral’s API on European, US, and global endpoints, or self-deployed — with customer control over where audio is processed and whether it is retained.

Mistral AI launched Voxtral 3 on October 20, and it is not just another model drop: it is the Paris lab’s bid to own the full voice-agent stack. The release bundles new speech-to-text and text-to-speech models with a voice agent SDK that fuses listening, reasoning, tool calls, and speaking into a single real-time streaming pipeline — the part of voice AI that has always been the hardest to get right. Mistral’s announcement frames it bluntly: building end-to-end voice agents that can listen, reason, interact, and handle complex workflows “remains difficult.” Voxtral 3 is its answer.

Voice has quietly become the most contested front in the model wars. Mistral arrives with a distinct pitch: speed (sub-second turns), European hosting and self-deployment for data-sovereign customers, and prices that undercut the established US APIs.

The models: ears first, then the voice

On the listening side, Voxtral Small Transcribe 3 is the flagship transcriber, covering 50+ languages including the 24 official EU languages plus low-resource Asian and Indic languages. It ships with speaker diarization (who said what), context biasing — feeding the model key terms or free-form text so it catches domain vocabulary without fine-tuning — and end-of-turn detection so agents reply without awkward pauses. Mistral claims state-of-the-art results in diarization and multilingual transcription, including in English, French, and Spanish, and says it outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe. The headline number: a 4.1% macro-average word error rate on the FLEURS 45 multilingual benchmark when the language is specified, 4.0% when it is inferred.

Two smaller siblings sit alongside it. Voxtral Mini Transcribe 3 is a 4-billion-parameter model for fast, accurate batch transcription at $0.0035 per minute. Voxtral Mini Transcribe Realtime 3 is built for agentic loops: sub-500ms latency across 35+ languages, with diarization and context biasing intact, at $0.008 per minute.

On the speaking side, Voxtral Speech 3 (the TTS side of the family, also called Voxtral Mini TTS 3) covers 35+ languages and 50+ dialects, adjusting emotion, pace, and emphasis to the conversation rather than reading flat. It adds voice cloning from 30+ minutes of sample audio, explicit instruction following for emotion, pacing, and emphasis, and a vocabulary library for tricky pronunciations. At $13 per million characters, Mistral positions it as cheaper to run than OpenAI’s TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. A Pro Voice tier offers higher-fidelity adaptation — reproducing a speaker’s distinctive character, emotional range, and speaking style across languages from 30 minutes of audio — for professional deployments where basic cloning falls short.

The SDK: one pipeline, sub-second turns

The voice agent SDK, built on LiveKit, is the centerpiece. The loop works like this: Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking; a general model — Mistral suggests its own Mistral Small — reasons and calls tools; Voxtral Speech 3 replies in the tone of the conversation so far. No stage waits for the one before it, so each turn completes in under one second, and the agent stops talking when the caller cuts in.

Crucially, the agent is not just talking — it can act mid-call. Mistral says agents can verify a caller’s identity, check a bill, or book an appointment using function calling, web search, and retrieval over the customer’s own systems. Agents answer and place calls over SIP, run in browsers and apps over WebRTC, and plug into the telephony, IVR, and contact-center software companies already run. Each stage is a separate model, so you can swap the LLM, the voice, or the language without rewriting the agent.

Pricing and the sovereignty pitch

Mistral’s full price list: Mini Transcribe 3 (batch) at $0.0035/min, Small Transcribe 3 at $0.0040/min, Realtime at $0.008/min, Speech 3 at $13 per million characters, and the voice agent API at $0.04/min with Mistral Small 4 doing the reasoning (or $0.08/min with any other reasoning model).

The other half of the pitch is deployment. Voxtral 3 is available through Mistral’s API on European, US, and global endpoints — or self-deployed on your own infrastructure. Organizations keep control over where conversations are processed, how data is handled, and whether any audio is retained. That sovereignty framing is the explicit wedge against US-hosted voice APIs, and it matches the pattern Mistral has run all year: open weights where it can, EU hosting where it matters.

A caveat, as with every launch: the accuracy and quality-per-dollar claims come from Mistral’s own benchmarks, including FLEURS runs and internal evaluations judged by Gemini 3.8 Flash, reported in the launch post and covered the same day by independent outlets. Independent verification will take time — but the stack is live today in Mistral’s API and Studio playgrounds, which is more than a benchmark table.