Mistral Voxtral 3 adds voice agents from $0.0035 a minute

Mistral AI released Voxtral 3 on October 20, 2026, a family of speech models for building voice agents that listen, reason, talk and act. Transcription starts at $0.0035 a minute and a full voice agent at $0.04 a minute. Mistral also claims the models beat closed rivals from OpenAI and ElevenLabs, Crypto Briefing reports.

What is in Voxtral 3

Voxtral 3 covers both directions of speech. Speech-to-text models turn audio into text, and a text-to-speech model turns text back into a voice. A voice agent API joins them to a language model, so a caller can talk to software that answers and takes actions.

Mistral's announcement lists these models and prices:

Model	What it does	Price
Voxtral Mini Transcribe 3	Batch speech-to-text	$0.0035 a minute
Voxtral Small Transcribe 3	Speech-to-text with speaker labels	$0.0040 a minute
Voxtral Mini Transcribe Realtime 3	Live speech-to-text	$0.008 a minute
Voxtral Speech 3	Text-to-speech	$13 per million characters
Voice agent API, with Mistral Small 4	Full voice agent	$0.04 a minute
Voice agent API, with another model	Full voice agent	$0.08 a minute

The post names the voice model several ways, including Voxtral Mini TTS 3. Transcription covers more than 50 languages, including the 24 official languages of the European Union. The voice model covers more than 35 languages and 50 dialects, with ready-made voices in 18 languages. A Pro Voice option builds a custom voice from 30 minutes of recorded speech.

How the voice agents work

The agent kit is built on LiveKit, an open-source toolkit for live audio and video. It connects to phone lines over SIP and to browsers and apps over WebRTC. Each stage is a separate model, so a team can swap the language model, voice or language without rewriting the agent, Mistral says.

Speed is the selling point. Mistral says live transcription runs in under 500 milliseconds, and a full agent turn finishes in under one second. Agents can call outside systems mid-conversation, The Decoder reports, for jobs like booking appointments or checking invoices.

The models run on Mistral's API, with endpoints in Europe, the US and globally. Companies can also deploy them on their own servers or private cloud by contacting Mistral.

Benchmarks and open questions

Mistral's numbers are its own tests. On speaker-attributed transcription, where the model must also say who spoke, it reports an error rate of 18.4% for Voxtral Small Transcribe 3. That beats Scribe v2 at 23.0% and Gemini 3.5 Transcribe at 38.4%, as the chart above shows.

On catching key terms in earnings calls, given a list of expected words, Voxtral Small Transcribe 3 scored 94.6%. Scribe v2 scored 94.3% and Gemini 3.5 Transcribe 87.3%.

Two reports conflict with Mistral's own page. Crypto Briefing describes the pipeline as open-weight and quotes about $0.001 a minute. Mistral's post states no license and lists $0.0035 to $0.008 a minute for transcription. Until Mistral publishes weights and a license, treat Voxtral 3 as an API product with a self-hosting option.

How it compares on price

Voice agent pricing is converging. xAI's Grok Voice 3.0 costs $0.08 a minute in its voice agent API. Mistral matches that with any outside model and halves it with its own Mistral Small 4. Voxtral 3 is Mistral's fourth Voxtral release since July 2025, after speech models in February and March 2026.

What this means for developers

Voxtral 3 makes a full voice agent a priced, off-the-shelf product rather than a stitched-together project.

• Price your call volume per minute. At $0.04 a minute, a 10,000-minute month costs about $400 with Mistral Small 4. Run your own call mix before comparing vendors.
• Test speaker labels on your audio. The 18.4% error rate comes from Mistral's chosen test sets. Record a few real calls with crosstalk and check who-said-what yourself.
• Use the modular design. Because each stage is separate, you can keep Mistral's transcription and plug in another language model. That costs $0.08 a minute instead of $0.04.
• Ask about weights before planning self-hosting. On-premises deployment goes through Mistral's sales team, and no license is published. Get terms in writing if data must stay in-house.