Mistral Voxtral 3: A Voice Agent SDK, New Speech Models and the Prices Mistral Published

Mistral released Voxtral 3 on October 22: a family of transcription and text-to-speech models, plus a voice agent SDK built on LiveKit. The pitch is one streaming pipeline that listens, reasons, calls tools and speaks, with each turn finishing in under a second. Those are Mistral's claims, and I have not run any of it.

I read one primary source: Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages. The benchmark charts are images and interactive plots, so I could not read the scores. I only have the captions and the text. I did not read the API docs.

What is in the release

Model or product	What Mistral says it is
Voxtral Small Transcribe 3	Speech-to-text across 50+ languages, with diarization, context biasing and end-of-turn detection
Voxtral Mini Transcribe 3	4B-parameter batch transcription
Voxtral Mini Transcribe Realtime 3	Realtime transcription with sub-500 ms latency, built for agents
Voxtral Mini TTS 3	Text-to-speech with emotion, pace and emphasis control, 35+ languages and 50+ dialects
Voice agent SDK	Listening, reasoning, tool calling and speaking in one real-time pipeline, built on LiveKit
Pro Voice	Voice adaptation from about 30 minutes of audio

Published prices

Mistral lists these prices on the launch page:

Item	Price
Voxtral Mini Transcribe 3 (batch)	$0.0035 per minute
Voxtral Mini Transcribe Realtime 3	$0.008 per minute
Voxtral Small Transcribe 3	$0.0040 per minute
Voxtral Mini TTS 3	$13 per million characters
Voice agent API with Mistral Small 4	$0.04 per minute
Voice agent API with a larger model	$0.08 per minute

Mistral also says the TTS costs less to run than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. I did not check those vendors' prices, and the comparison is Mistral's.

The voice agent SDK

Per the post, the pipeline has three stages. Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished. A general model such as Mistral Small reasons and calls tools. Voxtral TTS replies in the tone of the conversation. Mistral says no stage waits for the one before it, each turn completes in under one second, and the agent stops talking when the caller cuts in.

It says agents can verify a caller, check a bill or book an appointment through function calling, web search and retrieval over your own systems. It also says the agent can answer and place calls over SIP, run in browsers and apps over WebRTC, and connect to telephony, IVR and contact-center software. Each stage is a separate model, so you can swap the LLM, the voice or the language.

Where and how to run it

The post says Voxtral 3 is available through the Mistral API on European, US and global endpoints, or for self-deployment, and frames this as data control: where audio is processed and whether conversations are retained. The summary at the top says "EU-hosted API". The later text lists three endpoints. Check which region your data actually goes to. I did not find self-hosting requirements or licence terms in what I read.

Reading the benchmark claims

Mistral says Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe, including in English, French and Spanish. The captions show what the tests are:

• Multilingual transcription uses FLEURS, over the 45 languages all seven models support.
• Diarization uses public sets including AMI, Switchboard and AISHELL-4, with 30-minute versions of AMI and AISHELL for compatibility.
• Keyphrase recall uses Earnings-22 with 0 to 100 distractor terms.
• The TTS comparisons are mostly internal evaluations. Judging is by a Gemini model, on 300 English prompts for pronunciation and the numbers-and-names test, and the captions say prices come from vendor pages and results were generated on October 20.

An LLM judge and internal prompt sets are useful signals, not neutral measurements. One of the tests, Voice Code Bench, is credited to BeSimple AI. I could not read the scores themselves.

An inconsistency to note

The page describes the realtime transcription model as covering "over 35 languages" in one place and "40+ languages" in another. The FLEURS realtime chart uses 35 languages. Check the docs for the real list before you rely on a language.

What I would test

These are my own untested suggestions.

1. Test your own audio. Use real calls with your accents, noise and domain terms, and compare word error rate against your current provider.

2. Measure end-to-end latency from the caller's last word to the first audio, on your telephony path, not just the model figure.

3. Test interruptions. Check that the agent stops speaking when cut in, and that it does not cut people off.

4. Check numbers and names. Read back phone numbers, amounts, dates and emails. Mistral's own examples target these failure cases, so test them hard.

5. Price a full call. Add transcription, the reasoning model and TTS against the per-minute agent price.

6. Check consent and retention. If you clone a voice or record calls, look at the data and consent terms for your jurisdiction.

What I would not conclude

• That it beats Whisper or Gemini on your audio. The post's comparison is Mistral's and I could not read the scores.
• That every turn takes under one second in production.
• That the price is the total cost of a voice agent.
• That all 50+ languages are equally good.

Frequently asked questions

What is Voxtral 3?

A Mistral family of speech-to-text and text-to-speech models plus a voice agent SDK, announced October 22.

How much does it cost?

Mistral lists $0.0035 to $0.008 per minute for transcription, $13 per million characters for TTS, and $0.04 or $0.08 per minute for the voice agent API.

What is the voice agent SDK built on?

LiveKit, per the post.

Can I self-host it?

The post says it is available for self-deployment. I did not find hardware needs or licence terms.

How many languages does it support?

Transcription is listed at 50+ languages, realtime at 35+ or 40+ depending on where you read, and TTS at 35+ languages and 50+ dialects.

Can it clone voices?

Mistral describes voice cloning from 30+ minutes of audio and a Pro Voice option from about 30 minutes.

Is the speed claim verified?

No. Under one second per turn and sub-500 ms realtime transcription are Mistral's figures.

Sources

• Mistral AI, Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages (October 22, 2026): models, prices, SDK description, availability and benchmark captions. Chart scores were not readable.