October 20, 2026 — Mistral AI has launched Voxtral 3, a family of speech models plus a voice agent SDK that combines listening, reasoning, tool calling, and speaking in one real-time streaming pipeline. Why it matters: Mistral is selling a complete, self-hostable voice agent stack — with EU data control — at prices that undercut the US voice incumbents.
What happened
Voxtral 3, announced October 20, has three building blocks. Voxtral Small Transcribe 3 handles speech-to-text across more than 50 languages, including all 24 official EU languages. Voxtral Speech 3 (the text-to-speech model, also called Voxtral Mini TTS 3) speaks in more than 35 languages and 50-plus dialects, adjusting emotion, pace, and emphasis to the conversation. And a new voice agent SDK, built on LiveKit, ties transcription, a reasoning model, tool calls, and spoken replies into a single streaming system.
The details
The transcription models identify who said what (diarization), accept keyphrases and context to get specialized vocabulary right without fine-tuning, and detect when a caller has finished speaking so the agent can respond without delay. A realtime transcription variant runs at under 500 milliseconds of latency in more than 35 languages. Mistral says its Small transcription model beats Whisper, Scribe v2, and Gemini 3.5 Transcribe in its multilingual benchmarks — those are Mistral’s own published comparisons, not independent tests.
During a live call, an agent built on the SDK can verify a caller’s identity, check a bill, or book an appointment using function calling, web search, and retrieval over a company’s own systems, over phone (SIP) or web (WebRTC). Mistral says each turn completes in under a second, and the agent stops talking when the caller interrupts.
A feature called Pro Voice adapts a speaker’s voice — character, emotional range, and style — from 30 minutes of reference audio, for professional voice cloning.
Pricing is aggressive: batch transcription starts at $0.0035 per minute (Voxtral Mini Transcribe 3, a 4-billion-parameter model), realtime transcription is $0.008 per minute, speech generation is $13 per million characters, and the full voice agent API costs $0.04 per minute with Mistral Small 4 doing the reasoning, or $0.08 per minute with another reasoning model. The whole stack is available today through Mistral’s API — hosted in Europe, the US, or globally — or self-deployed on a customer’s own infrastructure.
Why it matters
Voice agents are where AI labs are now competing on complete products rather than raw models, and Mistral’s pitch is sovereignty plus price: run the entire stack in Europe or on your own servers, control whether audio is retained, and pay a fraction of a cent per minute for transcription. That combination is aimed squarely at European contact centers and regulated industries that cannot send call audio to US clouds. The benchmark claims still need independent testing, but a sub-second, tool-using voice agent that a company can host itself is a genuinely different offer from API-only rivals.
Frequently asked questions
What is Voxtral 3?
Voxtral 3 is Mistral AI’s voice family: speech-to-text models covering 50+ languages, a text-to-speech model covering 35+ languages and 50+ dialects, and a LiveKit-based SDK that combines listening, reasoning, tool calling, and speaking into one real-time voice agent pipeline.
How much does Voxtral 3 cost?
Transcription starts at $0.0035 per minute for batch and $0.008 per minute for realtime. Speech generation costs $13 per million characters. The voice agent API is $0.04 per minute with Mistral Small 4 for reasoning, or $0.08 per minute with another model.
Can Voxtral 3 run on a company’s own servers?
Yes. Mistral offers the stack through its EU-, US-, and globally hosted APIs, or as a self-deployment on the customer’s own infrastructure, with control over where audio is processed and whether it is retained.
How many languages does Voxtral 3 support?
Transcription covers more than 50 languages; speech output covers more than 35 languages and more than 50 dialects, with voice presets in 18 languages.
Sources: Mistral AI, The Decoder, TechnoBezz