On 20 October 2026, Mistral released Voxtral 3, a family of speech models plus a software kit for building voice agents: assistants that listen, think and answer out loud in real time, on the phone or in an app. It arrives two weeks after the company's Mistral Large 4, and it leans on the same pitch: competitive quality, low prices and hosting in the EU for teams with data residency rules.
The Voxtral 3 family
Small Transcribe 3
Batch, best accuracy- Price
- $0.004 per minute
- Speaker labels (diarization) and custom vocabulary
- 2.1% word error rate on Artificial Analysis, ranked third
For meetings, interviews and call recordings.
Mini Transcribe 3
Batch, cheapest- Price
- $0.0035 per minute
- Size
- 4 billion parameters
- Almost as accurate as Small (2.2%)
For large archives of audio.
Realtime Transcribe 3
Streaming- Price
- $0.008 per minute
- Detects when a speaker has finished, for natural turn-taking
- Less accurate than Google's real-time model in Mistral's own comparison
For live captions and voice agents.
Speech 3
Text-to-speech- Price
- $13 per million characters
- Over 35 languages and 50 dialects, with emotional tone
For agents that answer out loud.
How accurate is it?
Lower is better. Independent test; Mistral's models ranked third (Small) and close behind (Mini).
Lower is better. Mistral-reported comparison on FLEURS.
For recorded audio, Voxtral is among the most accurate models available, at one of the lowest prices. For live speech it is fast but makes noticeably more mistakes than Google's best, and Microsoft's MAI-Transcribe-2 processes batch audio several times faster. If you need transcripts of recordings, Voxtral is an easy choice; for live captions where every word matters, test it against the alternatives.
The voice agent kit
The SDK handles the plumbing that makes voice agents hard: streaming audio, deciding when the caller has finished, interruptions and phone connections. Agents can call your systems mid-conversation, for example to check an invoice or book an appointment. Running the whole agent through Mistral costs $0.04 per minute with Mistral Small 4, or $0.08 with other reasoning models.
Turning calls into data
For analytics teams, the interesting part is not the voice but the transcripts. At about 21 cents per hour of audio, transcribing every sales call or support line becomes cheap enough to do by default, and the text can then be searched, classified and counted like any other dataset.
- 1
Transcribe with speaker labels
Use Small Transcribe 3 with diarization, so you know who said what.
- 2
Store as rows
One row per speaker turn: call ID, speaker, start time, text.
- 3
Classify with a small model
Tag topic, sentiment and outcome with a cheap model such as Claude Haiku 5.5 or Mercury 3.
- 4
Report like any other data
Count complaints by topic per week, or link call outcomes to sales in your warehouse.
Small Transcribe 3Mini Transcribe 3
List prices in USD: $0.004 and $0.0035 per minute.
Check your understanding
1/3What does Voxtral Mini Transcribe 3 cost?
$13 per million characters is the text-to-speech price.
2/3Where does Voxtral trail its rivals?
At a 240 ms delay, Gemini 3.5 Realtime made fewer errors in Mistral's comparison.
3/3What is the voice agent SDK built on?
It handles phone (SIP) and app (WebRTC) sessions.
You got 0 out of 3.
More from Mistral: Mistral Large 4 explained. For more speech and agent news, see agents that work for hours.
Sources: Mistral's Voxtral 3 announcement as reported by MarkTechPost and The Decoder (20 October 2026); Artificial Analysis speech-to-text benchmark. Except for the Artificial Analysis results, accuracy figures are Mistral's own.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me