tl;dr: Mistral has released Voxtral 3, a family of speech-to-text and text-to-speech models paired with a LiveKit-based voice agent SDK that unifies listening, reasoning, tool-calling, and speaking into a sub-second streaming pipeline. The models claim state-of-the-art multilingual transcription (50+ languages), diarization, and context biasing, beating Whisper, Gemini 3.5 Transcribe, and Scribe v2 on benchmarks, with TTS supporting 35+ languages and voice cloning from 30 minutes of audio. Pricing starts at $0.0035/min for batch transcription and $13/1M chars for TTS, available via EU-hosted API or self-deployment.