The French AI lab is bundling transcription, text-to-speech, and a voice agents API into one open-weight pipeline aimed at OpenAI and ElevenLabs
Mistral AI announced Voxtral 3 on October 20, 2026, a suite of speech-to-text, text-to-speech, and voice agent models built for real-time conversation.
What Voxtral 3 actually does #
Voxtral 3 has three core pieces: Transcribe for turning speech into text, a text-to-speech (TTS) engine for turning text back into speech, and a new voice agents API with an accompanying SDK.
The voice agents API stitches together real-time transcription, reasoning models that process what was said, and streamed spoken replies. Mistral describes the result as a complete pipeline for natural voice conversations.
The company is also aiming the toolkit at workflow automation, not just chit-chat. That means voice agents that can kick off tasks, not only answer questions.
On language coverage, Mistral’s own summary says the models support over 50 languages. A more detailed breakdown counts more than 35 languages and more than 50 dialects.
Pro Voice, pricing, and the benchmark claims #
One of the more eye-catching additions is a feature called Pro Voice. It allows high-fidelity voice adaptation using up to 30 minutes of reference audio.
API access is expected to cost approximately $0.001 per minute. At that rate, a full hour of processing would run about six cents.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Mistral also claims Voxtral 3 outperforms established closed models from OpenAI and ElevenLabs on various benchmarks, pointing specifically to transcription accuracy and how expressively the models handle language.
It further claims Voxtral 3 can beat Whisper large-v3, OpenAI’s well-known speech recognition model, on various benchmarks.
How Mistral got here #
Voxtral 3 is not Mistral’s first attempt at speech. The company launched its initial speech-understanding models in July 2025, followed by TTS enhancements in March 2026.
Mistral has leaned on open-weight models, meaning developers can access the underlying model weights rather than renting a black box through an API. That stands in contrast to the proprietary approach favored by OpenAI and ElevenLabs.
What this means for developers and the voice AI market #
A low per-minute price combined with an all-in-one SDK removes two of the biggest hurdles for developers building voice features: cost and integration complexity. Until now, building a capable voice agent often meant combining a transcription service from one vendor, a language model from another, and a voice engine from a third.
The open-weight angle matters for companies handling sensitive audio, as running models on their own infrastructure can be a deciding factor for industries where sending customer calls to a third-party cloud raises compliance questions.
There are open questions worth watching. The first is whether independent evaluations confirm Mistral’s claims against OpenAI, ElevenLabs, and Whisper large-v3. The second is how the expected pricing of approximately $0.001 per minute holds up once real-world usage scales. The third is how Mistral manages consent and misuse around the Pro Voice cloning capability, given that high-fidelity voice adaptation from 30 minutes of audio is a powerful feature. Finally, support for more than 35 languages and over 50 dialects is broad, but businesses serving specific regional markets will want to test performance in their own target dialects before committing.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our