Mistral Releases Voxtral 3: Transcription in 50+ Languages, Speech at $13 per Million Characters and a Voice-Agent SDK

Mistral, the French company behind Le Chat, released Voxtral 3 on October 20: a set of speech-to-text and text-to-speech models plus an SDK for building voice agents that listen, reason, speak and act.

The models and prices

Voxtral Small Transcribe 3: the most accurate transcription model, $0.004 per minute.
Voxtral Mini Transcribe 3: batch transcription, 4 billion parameters, $0.0035 per minute.
Voxtral Mini Transcribe Realtime 3: streaming transcription for agents, $0.008 per minute, with latency under 500 ms.
Voxtral Mini TTS 3: text-to-speech in batch and realtime, $13 per million characters.
Voice agent SDK: built on LiveKit, $0.04 per minute with Mistral Small 4 doing the reasoning or $0.08 with a larger model. Mistral says each agent turn completes in under a second.

Languages and voices

The Transcribe 3 models cover more than 50 languages, including all 24 official EU languages. The speech model covers 35+ languages and 50+ dialects, takes instructions on emotion, pacing and emphasis, and offers voice presets in 18 languages. It also supports voice cloning, and a Pro Voice option adapts a voice from 30 minutes of a speaker's audio.

Mistral's own benchmarks put Voxtral Small Transcribe 3 ahead on speaker-attributed transcription: an error rate of 18.4%, against 23.0% for ElevenLabs Scribe v2 and 38.4% for Gemini 3.5 Transcribe. Mistral also says Voxtral Small beats Whisper on the multilingual FLEURS test and that its speech model is cheaper to run than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. These are vendor numbers, not independent tests.

Where it runs

Voxtral 3 is available now through Mistral's API, hosted in the EU with US and global endpoints, and in Mistral Studio playgrounds for testing transcription, speech, voice cloning and agents. Companies can also deploy it on their own infrastructure; on-premises and private-cloud setups go through Mistral's sales team.

Who it matters to

The release targets developers building call-centre bots, meeting transcription and voice assistants, especially those who need audio to stay in the EU or on their own servers.

What this means if you need speech AI

If you transcribe meetings or calls in several languages, Voxtral Mini at $0.0035 a minute is worth a test against the open Whisper model you may already run.

If you need natural voices for content, ElevenLabs remains the ready-made studio; Voxtral is an API for developers, so compare it on your own scripts before switching.