Mistral promises voice agents that can answer the phone in under a second

Under one second. That is how long it takes, according to Mistral AI, for a voice agent built with Voxtral 3 to hear a question, understand it, query a system and answer out loud. The French start-up unveiled the new family of voice models on October 20, along with a development kit for building assistants that can pick up a phone call.

The offering is available now through Mistral's API, with European, U.S. and global endpoints. It includes three transcription models and one text-to-speech model. According to the company, transcription covers more than 50 languages, including the 24 official languages of the European Union, and speech synthesis covers more than 35 languages.

What's in the Voxtral 3 family?

• Voxtral Small Transcribe 3, billed as the most accurate, at $0.004 per minute;
• Voxtral Mini Transcribe 3, a 4-billion-parameter model for transcribing files, at $0.0035 per minute;
• Voxtral Mini Transcribe Realtime 3, for live transcription with latency under 500 milliseconds according to Mistral, at $0.008 per minute;
• Voxtral Mini TTS 3, which turns text into speech, at $13 per million characters.

The transcription can tell who is speaking, and it accepts a list of business terms to better recognize technical vocabulary, without retraining. The agent kit, built on the open-source software LiveKit, chains listening, reasoning and response. It plugs into telephony (SIP) and browsers (WebRTC), stops talking when the person on the line interrupts, and can call tools or query internal documents. With the Mistral Small 4 model as its "brain," the agent costs $0.04 per minute.

Performance that others need to confirm

Mistral claims its transcription model outperforms Whisper (OpenAI), Scribe v2 (ElevenLabs) and Gemini 3.5 Transcribe (Google). On the multilingual FLEURS benchmark, its charts show a 4.1% word error rate for Voxtral Small Transcribe 3, against 16% for Whisper large-v3. But Mistral ran these measurements itself. For speech synthesis, only one test was entrusted to an outside firm, BeSimple AI, and it was scored by a Google AI model. The rest comes from internal evaluations. No independent evaluation was available at the time of writing.

Another unknown is the license. The first Voxtral models, in July 2025, were released with their weights under the open Apache 2.0 license. The Voxtral 3 page says nothing of the kind. It only mentions deployment on the customer's own infrastructure, on request through the sales team.

What to check before you start

For a call center or customer service team, the appeal is clear: a European provider, possible hosting within the EU and control over how recordings are retained, according to Mistral. A few points still deserve attention:

• Voice cloning: Mistral offers to reproduce a voice from about thirty minutes of recording. The page describes no safeguards against abuse, even though impersonating an executive's voice is already used in wire-transfer scams.
• Voice data: a voice is personal data. You need to inform callers, set a retention period and check where recordings are processed.
• Testing: run tests on your own recordings, with your accents and your jargon, rather than relying on the vendor's charts.

After Timestral for forecasting and Mistral Large 4, the start-up is rolling out specialized models one after another this October. The next announced step: the release of Large 4's weights, which Mistral has promised by the end of the month. For Voxtral 3, the license question remains open.