Mistral Voxtral 3: Mistral’s speech models for transcription, TTS and agents

Batch transcription costs $0.0035 to $0.004 per minute, and TTS costs $13 per 1M characters. The realtime transcription model is priced at $0.008 for the realtime model. The TTS model reads the meaning of the text to pick emotion, pace and emphasis. Mistral lists voice cloning from roughly 30 minutes of audio. The LiveKit-based SDK runs transcription, reasoning and speech output as one streamed sequence, and each turn completes in under one second. The company launched its initial speech-understanding models in July 2025, followed by TTS enhancements in March 2026. Mistral positions the family against OpenAI, Google, ElevenLabs, Deepgram, and Microsoft. Its performance claims are vendor-reported, and no independent replication yet exists. Coverage figures also need checking: the TTS documentation supports 9 languages, while launch material cites 35+ languages. For a team building an AI company, the pricing is the main signal. A cascaded voice agent lands in the low cents per minute before telephony. Deployment options matter too, since the models are available via Mistral's EU-hosted API or self-deployment. Open weights were not announced for the realtime model, so teams needing self-hosting should confirm terms before committing. Test on your hardest audio, including accents and background noise, before switching providers.