Source: Mistral AI news
Mistral releases Voxtral 3 speech models and a Voice Agents API
Why it matters
Four Voxtral 3 models are in Public Preview on the Mistral API: Voxtral Mini Transcribe 3 ($0.0035/min), Voxtral Small Transcribe 3 ($0.004/min), Voxtral Mini Transcribe Realtime 3 ($0.008/min, configurable latency of 80 to 2,400 ms with end-of-turn detection) and Voxtral TTS 3 ($13 per 1M characters, down from $16 for the previous Voxtral TTS).
Voice Agents is a managed pipeline (transcription, model, speech, playback, interruptions, tool calls) priced at $0.04/min with Mistral Small 4 or $0.08/min with larger models; install with pip install "mistralai[voice-agents]" plus PortAudio and start a session with client.audio.sessions.init(), in Python, Node.js or the browser.
Mistral says transcription covers 50+ languages; its TTS 3 model card lists 9 languages while the announcement says 35+, so check your language before you build. Limits: all four models are preview, the model cards are marked Premier and no Voxtral 3 weights are on Mistral's Hugging Face page yet, phone (SIP) agents are enterprise-only at launch, and Mistral says Small Transcribe 3 beats Whisper, Scribe v2 and Gemini 3.5 Transcribe, on benchmark charts it published itself.