French AI lab Mistral AI releases Voxtral 3 voice model suite
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
French AI lab Mistral AI released Voxtral 3, a speech-to-text, text-to-speech and voice agent model suite.
It targets contact centers and assistants, and Mistral claims it beats OpenAI and ElevenLabs.
Background
Mistral AI is a French AI lab. Voxtral 3 is its latest speech stack: a family of speech-to-text and text-to-speech models plus a voice agent SDK (a toolkit for building voice apps) that combines listening, reasoning, tool calling and speaking in one real-time streaming pipeline. It is not Mistral's first speech attempt — the company launched initial speech-understanding models in July 2025 and added text-to-speech enhancements in March 2026. Mistral has leaned on open-weight models, letting developers access the underlying model weights rather than renting a black box through an API, in contrast to the proprietary approach favored by OpenAI and ElevenLabs.
How it unfolded
- Jul 2025Mistral launched its initial speech-understanding models.
- Mar 2026Mistral followed with text-to-speech enhancements.
- Oct 20, 2026Mistral AI announced Voxtral 3, a suite of speech-to-text, text-to-speech and voice agent models built for real-time conversation.
- Oct 20, 2026Voxtral 3 became available through Mistral's EU-hosted API or for self-deployment on an organization's own infrastructure.
Who’s saying what
- Official
- Mistral says Voxtral 3 is a family of state-of-the-art speech models plus a voice agent SDK, with its realtime model delivering transcriptions at latency below 500 milliseconds.
- Official
- Mistral says Voxtral 3 outperforms established closed models from OpenAI and ElevenLabs on benchmarks for transcription accuracy and expressiveness.
Still unverified
Mistral's claims that Voxtral 3 beats Whisper large-v3 and closed models from OpenAI and ElevenLabs on various benchmarks are attributed to Mistral itself. Pricing figures differ across sources: $0.0035 per minute for transcription versus approximately $0.001 per minute for API access.