Mistral prices a full voice agent at $0.04 a minute

Voice agents just got a price you can put in a spreadsheet.

Mistral released Voxtral 3 on Oct. 20: a set of transcription and text-to-speech models and a voice agent API that strings them together with a Mistral model doing the reasoning. The company prices the whole agent by the minute.

$0.04 a minute, with Mistral Small 4 doing the thinking.

Four models and one agent pipeline

Mistral lists these parts and prices:

• Voxtral Small Transcribe 3, its most accurate transcription model, at $0.0040 a minute, with 50+ languages, speaker diarization and context biasing for domain terms.
• Voxtral Mini Transcribe 3, a 4B-parameter batch model, at $0.0035 a minute.
• Voxtral Mini Transcribe Realtime 3, built for agents, at $0.008 a minute, with latency under 500 milliseconds, Mistral says.
• Voxtral Speech 3, the text-to-speech model, at $13 per million characters, with 35+ languages and 50+ dialects.

The voice agent API is built on LiveKit. The realtime model listens, Mistral Small reasons and calls tools, and Voxtral Speech 3 talks back. Mistral says each turn completes in under a second and the agent stops talking when the caller interrupts. It works over SIP phone lines and over WebRTC in browsers and apps, with function calling, web search and retrieval over your own systems.

Swap in any other reasoning model and the price doubles to $0.08 a minute.

Mistral also claims a big accuracy lead. On its FLEURS test across 45 languages, it reports a 4.1% word error rate for Voxtral Small Transcribe 3 with the language set, against 16.0% for OpenAI's Whisper large-v3. Those are Mistral's numbers, on Mistral's chosen tests.

The price is the launch

Yes, every voice company has a chart where it wins. Treat Mistral's like the others until someone independent reruns it.

The price is different. Voice agents are usually sold as a bundle of separate bills: one vendor for speech to text, one for the model, one for the voice, often a fourth for the telephony glue. Mistral is quoting one number per minute for the stack, and publishing it.

The design helps too. Each stage is a separate model, so Mistral says you can swap the language model, the voice or the language without rebuilding the agent. That keeps you from being locked into one vendor's brain just because you like its ears.

And the deployment story is aimed at regulated buyers. The API runs on EU, US and global endpoints, and Mistral says you control where audio is processed and whether it is kept. Self-hosting is available, but only through Mistral's sales team.

What to check before you build on it

A few things the announcement does not settle:

• License. Mistral's post does not say whether these models ship with open weights, or under what terms. Do not assume they do.
• Real latency. "Under a second" per turn is Mistral's claim. Measure it on your own calls, from your region, with your tools in the loop.
• Language quality. Transcription covers 50+ languages, but voice presets exist in 18. If your callers speak a language outside those, test it first.
• Voice cloning. Mistral says it can clone a voice from 30 or more minutes of audio. Get consent in writing before you use anyone's voice.

If you already run a voice agent, do the math on one month of call minutes at $0.04 and compare it with your current bill, line by line.

What to watch

Watch for independent benchmarks of the realtime model, and for whether Mistral publishes a license for the weights. Then watch what the other voice API vendors charge by the end of the quarter.