Mistral releases Voxtral 3, speech models and a voice agent SDK

Mistral announced Voxtral 3 on October 20, 2026 (shared on Hacker News, where it had 735 points and 182 comments by October 20). Details and benchmarks below are Mistral's own claims from its announcement page.

• Speech-to-text: Voxtral Small Transcribe 3 and a 4B-parameter Voxtral Mini Transcribe 3 ($0.0035/min), covering 50+ languages with diarization, context biasing and end-of-turn detection. A Voxtral Realtime Transcribe 3 model offers sub-500 ms latency in 35+ languages. Mistral says the Small model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe on its charts, including an Artificial Analysis benchmark.
• Text-to-speech: Voxtral Mini TTS 3 (batch and realtime) with 35+ languages and 50+ dialects, emotion, pace and emphasis control, and pronunciation hints. Mistral prices it at $13 per 1M characters and says it costs less than comparable models from OpenAI, Deepgram, ElevenLabs and Gemini. The TTS comparisons are Mistral's internal evals, judged by a Gemini model.
• Voice agent SDK: one real-time pipeline for listening, reasoning, tool calling and speaking, for phone or web use.
• Pro Voice: voice adaptation from about 30 minutes of audio.
• Availability: Mistral's EU-hosted API, or self-deployment.

Independent benchmark. On October 22, Voice Arena published results for the two transcription models on its Diarization + ASR leaderboard, tested on 139 held-out conversations with human-verified references, a 0 ms collar and no preset speaker counts. Voxtral Mini Transcribe 3 ranked #2 of 9 models by diarization error rate (DER) at 40.4%, ahead of Voxtral Small Transcribe 3 at 41.1% (#4). Voice Arena says Mini leads on in-person conversations and Small is slightly better on online calls, that missed speech (mostly overlapping speakers) makes up 82% of Mini's DER, and that both models counted the speakers correctly in 83.5% of sessions. Mistral shared the result on October 28.