Mistral Voxtral 3: A Voice Agent Stack at 4 Cents a Minute

Mistral is now selling a complete phone-ready voice agent, listening, reasoning, tool calls and speech, for $0.04 a minute when the reasoning runs on its own Mistral Small 4. For teams that have been wiring separate transcription, LLM and text-to-speech vendors together for contact-center bots, Voxtral 3 turns a three-vendor integration into one API, with the option to run all of it on their own hardware.

What Happened

As The Decoder reports, Mistral has launched Voxtral 3, a family of models for real-time voice agents: a speech-to-text model, Voxtral Small Transcribe 3, a text-to-speech model, Voxtral Mini TTS 3, and a voice agent SDK built on LiveKit. Transcription covers more than 50 languages, and speech output covers more than 35 languages and 50 dialects.

Mistral’s announcement fills in the rest. Voxtral Realtime Transcribe 3 transcribes at sub-500ms latency, and end-of-turn detection uses acoustic and semantic signals to judge when a caller has finished speaking. In the SDK, transcription, a general model that reasons and calls tools, and speech output run as one streaming pipeline where no stage waits for the one before it. Mistral says each turn completes in under one second, and the agent stops talking when the caller cuts in. Agents answer and place calls over SIP, run in browsers and apps over WebRTC, and can verify a caller, check a bill or book an appointment through function calling, web search and retrieval.

Why It Matters

The pricing is the clearest signal. According to Mistral’s price list, batch transcription with Voxtral Mini Transcribe 3 costs $0.0035 a minute, realtime transcription $0.008 a minute, and Voxtral Speech 3 costs $13 per million characters. The voice agent API costs $0.04 a minute with Mistral Small 4 doing the reasoning, and $0.08 a minute with any other reasoning model, so the cheapest path keeps the whole call on Mistral.

Mistral also makes benchmark claims a buyer should test rather than accept: that Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe, and that its speech model costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Several of its text-to-speech comparisons are labeled internal evaluations judged by a Gemini model.

Stack Archive’s earlier comparison of GPT-Transcribe against ElevenLabs, Google and Mistral showed how quickly per-minute transcription prices fell. Voxtral 3 moves that price competition from single models to the whole agent.

What This Means for Agent Builders

The design choice that matters most is that each stage is a separate model. Mistral says teams can change the LLM, the voice or the language without rewriting the agent, and run the stack through its European, US or global API endpoints or on their own infrastructure. That makes Voxtral 3 a sovereignty play as much as a price play: an EU contact center can keep audio inside its own environment and decide whether any of it is retained.

Two features target enterprise brand teams. Pro Voice adapts a speaker’s voice from 30 minutes of audio, and context biasing lets the transcriber favor supplied key terms and names without fine-tuning, so domain vocabulary doesn’t require a custom model. The Voxtral documentation covers the APIs, and full on-premises deployment goes through Mistral’s sales team.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.