What happened
On October 20, 2026, Mistral published Voxtral 3, a family of speech-to-text and text-to-speech models and a voice-agent SDK. Voxtral Mini Transcribe 3 is 4 billion parameters at $0.0035 a minute. Voxtral Small Transcribe 3 is $0.0040 a minute, and Mistral says it covers more than 50 languages. The realtime Mini model is $0.008 a minute, and Mistral says it transcribes in under 500 milliseconds across more than 35 languages. Voxtral Mini TTS 3 is $13 per million characters, with more than 35 languages and more than 50 dialects. A voice agent that uses Mistral Small 4 for the reasoning step is $0.04 a minute. The same agent with a larger model is $0.08 a minute. Mistral says Small Transcribe outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe. Those comparisons are charts.
Why it matters
A voice agent was a pile of vendors. Mistral is selling the listen step, the speak step, and the loop as one price list, with the reasoning model swappable. Mini transcription at 4 billion parameters is the piece a team might self-host. The post says self-deployment is available and does not name a public weight file or a license. The accuracy claim stays Mistral's until a word-error rate is written in the prose or someone else reruns the test.
Mistral says the voice-agent SDK runs on LiveKit. The realtime transcriber listens, a separate model such as Mistral Small reasons and calls tools, and the text-to-speech model replies. Mistral says a turn completes in under one second.
Pro Voice adaptation is described as needing 30 minutes of audio. The text-to-speech price comparison with other vendors is Mistral's, generated October 20, 2026, and judged with gemini-3.8-flash on Mistral's account of the setup.