Mistral Launches Voxtral 3, a Full Voice Agent Stack

Mistral shipped Voxtral 3 on Tuesday, and it is not another speech model release. It is an entire voice stack in one box: a transcription model, a text-to-speech model, and a voice agent SDK that runs all of it as a single streaming pipeline. Mistral says each conversational turn now completes in under one second (official announcement).

For anyone who has wired a voice agent together in the last two years, that framing is the story.

What is actually in the box

The stack has three parts. Voxtral Small Transcribe 3 handles speech-to-text across 50+ languages, with speaker diarization, context biasing for domain vocabulary, and native end-of-turn prediction so an agent knows when the caller has stopped talking instead of waiting on a silence timer. Voxtral Mini TTS 3 covers text-to-speech in 35+ languages and 50+ dialects, adjusting emotion, pace and emphasis based on what was actually said.

The third piece is the new voice agent SDK, built on LiveKit. It runs listening, reasoning and speaking as one pipeline where no stage waits for the one before it, so the agent can interrupt and be interrupted. During a live call it can verify identity, check a bill or book an appointment through function calling over your own systems, and it speaks SIP, WebRTC and the telephony and contact-centre software teams already run.

The price is the argument

Mistral is pricing the whole stack below the incumbents. Batch transcription starts at $0.0035 per minute, realtime transcription at $0.008 per minute with sub-500ms latency, and Voxtral Speech 3 at $13 per million characters. Mistral claims that last figure undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash, and it published a benchmark chart to back it. Realtime transcription runs at $0.008 per minute.

On quality, Mistral reports 4.1% word error rate on the FLEURS 45 multilingual benchmark with the language specified, and 4.0% when it is auto-detected, against 16.0% and 14.9% for Whisper large-v3 (as AI Weekly summarised). Those are Mistral's own numbers, not third-party evaluations, so treat them as a claim rather than a verdict.

There is also a deployment angle: Voxtral 3 runs through Mistral's EU-hosted API or self-deployed on your own infrastructure, with control over whether audio is retained. For European contact centres, that sovereignty story has been the missing piece for years. Pro Voice adaptation can also reproduce a speaker's character across languages from 30 minutes of audio.

What it means for builders

Voice agents have been a three-vendor assembly problem: one API to transcribe, one model to reason, one service to speak, and a glue layer that adds latency with every hop. Collapsing that into a single streaming SDK removes a class of failure and a class of invoices. It is the same shift visible in how agents are starting to act on the phone at all, as Yelp's restaurant-calling agent showed this week. The trade-off is consolidation: the pipeline, the reasoning model and the voice all come from Mistral, and the benchmark claims that justify the switch are still the vendor's own.