· Translated from the Spanish original

Voice and AI

Mistral launches Voxtral 3 and charges for transcription from US$ 0.0035 per minute

The suite adds three transcription models and one voice model, with latency under 500 ms. The comparisons with Whisper are Mistral's own, unaudited.

What happened

  • Mistral AI introduced Voxtral 3 on October 20, 2026, a voice suite with three transcription models and one speech synthesis model.
  • The transcription models are Small Transcribe 3 and Mini Transcribe 3 (4 billion parameters), both batch, and Mini Transcribe Realtime 3, for streaming. The voice model is called Voxtral Speech 3.
  • Mistral claims transcription in more than 50 languages and real-time latency under 500 milliseconds. A full voice-agent turn, it says, takes under one second.
  • The published prices are US$ 0.0035 per minute (Mini), US$ 0.004 (Small) and US$ 0.008 (real time). Speech synthesis costs US$ 13 per million characters.

Why it matters

  • Transcription is the cheap part of working with voice, and Mistral prices it in fractions of a cent. An SME that records sales calls can transcribe 1,000 hours for about US$ 210 with the Mini model, calculated at the published rate.
  • Spanish and Portuguese appear in the language lists of the tests. The publication does not detail how it performs with Chilean or Rioplatense accents, and a September paper measured that error grows for certain groups after pruning audio models.
  • The consequence the source does not state: price per minute is no longer the obstacle to automating the call center. The obstacle becomes the human review of what is transcribed, which Mistral does not price.

The number

4.1% versus 16.0%. Word error rate of Voxtral Small against Whisper large-v3 on FLEURS, 45 languages, with the language specified. It is a Mistral measurement.

Context

Competition was already in this territory. Microsoft bundled transcription in 60 languages and two voice generation models in a single announcement. Mistral also compares against Scribe v2 and Gemini 3.5 Transcribe.

What’s next

  • The publication sets no date for new languages or price changes.
  • The models are on Mistral’s API, with endpoints in Europe, the United States and global, and in Studio.
  • To run them on your own infrastructure you must talk to sales. The page states neither a license nor open weights.

Bottom line

On September 10, OpenAI opened its voice model to developers for phones and agents. Mistral enters from the transcription side, with a cost that invites testing and a license that is still unknown.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes