Mistral launches Voxtral 3 voice models and an agent SDK

Mistral launched Voxtral 3, a family of models that change voice to text and text to voice with an SDK (a software kit) for voice agents. The kit joins listening, reasoning, tool use and speaking in one realtime flow. Transcription covers 50+ languages. Mistral says its Small model gives better scores than Whisper, Scribe v2 and Gemini 3.5 Transcribe, but the launch text gives no scores.

What is new

Voxtral 3 has these parts:

• Small Transcribe 3 names who spoke each sentence and uses terms that you give for precision.
• Mini TTS 3 gives voices that change emotion and pace with the context.
• Realtime Transcribe 3 predicts when a user has finished speaking.

The numbers

Mistral gives these figures:

• The Mini transcription model has 4B parameters and costs $0.0035 for each minute.
• Voice output costs $13 for 1M characters, less than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.
• Realtime transcription has latency of less than 500 ms in 35+ languages.

What is not known

All claims about results are from Mistral. The launch text gives no scores for precision and no test by a group that is not Mistral.

Insight: Voice vendors sell two different bets

ElevenLabs launched Eleven Realtime on 20 October. Each model is a complete voice agent, with an API that works as the OpenAI Realtime API. Mistral sells parts and an SDK, with EU hosting or use on servers of the buyer. This reads as one bet on easy change of vendor and one bet on control of where audio goes. Monitor which bet buyers of voice agent tools favor.