Mistral Voxtral 3 Posts 4.1% WER on FLEURS, Beating Whisper

Mistral says its new Voxtral Small Transcribe 3 hits 4.1% word error rate on the FLEURS 45 multilingual benchmark when the language is specified, and 4.0% when it is inferred, against 16.0% and 14.9% for Whisper large-v3. The claim sits in Mistral's launch post, which also names Gemini 3.5 Transcribe, Scribe v2 and GPT Transcribe as beaten on the same test.

The headline product is a voice agent SDK the company says "combines listening, reasoning, tool calling, and speaking into a real-time experience". Three transcription tiers sit under it: Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.004 per minute, and Voxtral Mini Transcribe Realtime 3 at $0.008 per minute with sub-500ms latency. On the live-agent loop, the post states that "each turn completes in under one second, and the agent stops talking when the caller cuts in."

The text-to-speech model, Voxtral Speech 3, is priced at $13 per million characters, covers 35+ languages and 50+ dialects, and clones a voice from 30+ minutes of audio with instruction following for emotion, pacing and emphasis. On price, Mistral writes that "Voxtral Speech 3 is positioned as more cost-effective than OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash."

Deployment is through Mistral's EU-hosted API or self-hosting on private infrastructure, pitched as "sovereign deployment" with customer control over audio retention. The launch post publishes no per-language FLEURS breakouts and no customer names.

The release lands a day after Mistral's Timestral time-series model, making it the eleventh story we have tracked from the lab in the last three months.