Mistral launches Voxtral 3 voice agent stack

Mistral AI has launched Voxtral 3, a full voice agent stack that listens, thinks and talks in real time. The Paris-based company has bundled a speech-to-text model, a text-to-speech model and a voice agent SDK into a single streaming pipeline — the components being Voxtral Small Transcribe 3, Voxtral Mini TTS 3, and an agent kit built on LiveKit.

The specification that matters here is not benchmark scores but latency and cost per minute. Transcription starts at $0.0035 per minute. The transcription model covers more than 50 languages; the speech output supports over 35 languages and 50 dialects with emotional adaptation. Agents can reach external systems mid-conversation — booking appointments, checking invoices — which is the difference between a voice toy and something a call center could actually deploy.

Mistral's pitch is aimed squarely at the compliance layer European buyers care about: the stack can run on customers' own infrastructure or through EU-hosted APIs, keeping data under their control. That positioning, plus open weights and per-minute pricing, is a direct attempt on the API subscription model that American providers sell into the same market. For Europe's service sector — customer support, bookings, administration — a domestic voice stack at that price point changes the procurement calculus, provided the quality holds up under real conversational load rather than demo conditions.

The open question is distribution. OpenAI and Google own the consumer channels through which voice AI reaches most people, and a superior per-minute price does not matter to a customer who never hears about it. Mistral's bet is that enterprises buying voice agents in volume shop on cost, sovereignty and integration — and that the interface layer, not the frontier model, is where the European vendor can actually win.

The transcription alone covers more than 50 languages. Whether anyone dials in depends on channels Mistral does not yet own.