Summary: Mistral announced Voxtral 3 on October 20, 2026: a family of speech-to-text and text-to-speech models plus a voice agent SDK that chains listening, reasoning, tool calling, and speaking into one real-time streaming pipeline. It covers 50+ languages and is available through Mistral’s EU-hosted API or self-deployment, which puts data-sovereignty decisions directly in the builder’s hands.
Why this is more than another speech model
The hard part of a voice agent has never been any single capability. As Mistral frames it, a reliable agent must do four things simultaneously: listen accurately across languages, accents, and interruptions; reason about intent in context; speak naturally with expression; and take multi-step actions in connected systems. Getting all four to cooperate in real time is where most projects stall.
Voxtral 3 attacks that integration problem directly. The SDK combines the four loops into one pipeline, so an agent can retrieve information, hit business systems, and complete tasks mid-conversation over phone or web. Mistral’s demo shows the pattern: a caller reads a French referral letter number, the agent calls a tool, translates the letter’s gist back in English, and books the appointment in the same turn.
What the transcribe side delivers
Voxtral Small Transcribe 3 covers 50 languages, including 24 official EU languages and a number of low-resource Asian and Indic languages. Mistral reports state-of-the-art results on multilingual transcription and diarization, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe on their benchmarks. The 4B Mini variant transcribes at $0.0035 per minute.
Two details matter for production use. Context biasing lets you feed in key terms and free-form text so the model recalls domain vocabulary without fine-tuning — useful for product names, drug names, or internal jargon. And Voxtral Realtime Transcribe 3 targets agentic use cases at sub-500ms latency across 35+ languages, with native end-of-turn prediction that uses both acoustic and semantic signals to avoid cutting users off prematurely.
This builds on the transcribe line we covered in Mistral Voxtral Transcribe 2, which brought open speech-to-text to on-device deployment. The trajectory is consistent: Mistral keeps shipping the speech layer and now wraps it in an agent harness.
The speech side, and pricing
Voxtral Mini TTS 3 supports 35+ languages and 50+ dialects, in both batch and realtime modes, at $13 per 1M characters — which Mistral says undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash at comparable quality. Rather than requiring explicit emotion instructions, the model interprets the text’s context to pick emotion, pacing, and emphasis, though explicit control over those variables is available.
There’s also a vocabulary library for pronunciation control (phoneme hints for rare or invented words), voice cloning from 30+ minutes of audio, and a Pro Voice tier that reproduces a speaker’s character and emotional range across languages from 30 minutes of samples.
The sovereignty angle
Like other Mistral releases, Voxtral 3 is built for sovereign deployment: EU-hosted API or self-hosted on your own infrastructure, with control over where audio is processed and whether any conversation is retained. For contact centers and healthcare-adjacent workloads in Europe, that retention switch may be the deciding factor regardless of benchmark scores.
What to weigh before adopting
The benchmark claims are Mistral’s own or from named third parties like BeSimple AI’s Voice Code Bench, several judged by gemini-3.8-flash — worth replicating against your own traffic before committing. The practical entry points are clear, though: if you’re already assembling voice agents from separate ASR, LLM, and TTS vendors, a single SDK with native end-of-turn handling and context biasing removes real glue code. Start with the transcription layer if multilingual accuracy is your bottleneck; start with the SDK if orchestration latency is.