On October 20 Mistral introduced Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that runs listening, reasoning, tool calling and speaking as one streaming pipeline. In its announcement, the company calls its Small transcription model state-of-the-art for multilingual transcription and speaker labelling, and says it outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe.
Transcription, speech and an agent SDK share one price list
Voxtral Small Transcribe 3 and Mini Transcribe 3 cover 50+ languages, with speaker labelling and context biasing, which lets a developer feed in domain terms to improve recall. Mini has 4B parameters and is priced at $0.0035 a minute for batch work; Small costs $0.0040 a minute, and a real-time Mini variant costs $0.008. Voxtral Speech 3 covers 35+ languages and 50+ dialects at $13 per million characters. The agent SDK is built on LiveKit and chains a real-time transcription model, a reasoning model and the speech model, with each turn completing in under one second, according to Mistral. The voice agent API costs $0.04 a minute when Mistral Small 4 does the reasoning and $0.08 with any other model.
The comparisons are the ones Mistral chose to publish, and the speech scores use a model as judge
Its transcription charts report word error rate on FLEURS and other sets, including an index from Artificial Analysis. The text-to-speech charts, which set the model against ElevenLabs, OpenAI, Deepgram and Gemini, carry the label internal eval, and their footnotes say a Gemini model acted as judge and that prices came from the vendors’ pricing pages. That makes them Mistral’s measurements under its own prompts, not an independent ranking.
The models run on Mistral’s servers, and on-premises use goes through its sales team
Voxtral 3 is available today through the Mistral API on European, US and global API regions. For on-premises or private-cloud deployment, the post points readers to Mistral’s team, and it mentions no weights download, license or hardware requirement. For readers who run models locally, that appears to be the practical limit: Mini has 4B parameters, but the announcement does not say whether it can be pulled onto a GPU of your own.