we_are_coded.by CODE · The world, decoded
БГ

Mistral · Voice

Mistral released Voxtral 3: speech recognition in 50 languages and synthesis in 35+, and a full voice agent starts at $0.04 a minute

Mistral announced Voxtral 3 on October 20: speech recognition in 50 languages (Bulgarian is among those tested), synthesis in 35+ languages and an SDK for voice agents. Transcription starts at $0.0035 a minute, a full agent at $0.04. The comparisons with Whisper and the others are Mistral's own.

Source
Mistral
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

A minute of transcription costs $0.0035. A minute of a full voice agent that listens, thinks and answers costs from $0.04 to $0.08. Those are the prices of Voxtral 3.

Fact

The facts: on October 20, 2026 (the page carries no visible date; this is the day from Mistral's feed), Mistral announced Voxtral 3, a new family of speech-to-text and text-to-speech models, together with a voice agent SDK built on LiveKit. Recognition comes as Small Transcribe 3 and Mini Transcribe 3 (4B parameters), plus Mini Transcribe Realtime 3; the supported language list is expanded to 50, including 24 official EU languages. Synthesis, Mini TTS 3, supports 35+ languages and 50+ dialects. Prices per Mistral's table: Mini Transcribe 3 (batch) is $0.0035 a minute, Small Transcribe 3 $0.0040, Mini Transcribe Realtime 3 $0.008; Mini TTS 3 is $13 per 1M characters; the agent API is $0.04 a minute with Small 4 for reasoning and $0.08 with a larger model. Voxtral 3 is available through Mistral's API, with European, US and global endpoints, or for self-deployment; for the latter the page describes neither weights nor a license. The quality claims are Mistral's: that Small leads in recognition and diarization and is ahead of Whisper, Scribe v2 and Gemini 3.5 Transcribe, and for synthesis Mistral points to the Voice Code Bench by BeSimple AI.

The math is simple. An hour of conversation through the agent API comes to $2.40 with Small 4 reasoning and $4.80 with a larger model, per Mistral's table.

Bulgarian is there: you can see it in the list of 45 languages Mistral measures recognition on. Whether the synthesis speaks Bulgarian, the page does not confirm.

Listening is under a cent. The agent starts at four cents.

The comparisons are Mistral's. They put Small ahead of Whisper, Scribe v2 and Gemini 3.5 Transcribe on FLEURS, and for synthesis Mistral cites the Voice Code Bench by BeSimple AI, a test of customer-support failures around numbers and unique names, run on 300 prompts, all of them in English.

The first check is a recording of your own: a noisy line, an accent, an order number dictated in Bulgarian. If Voxtral 3 catches it, the rest is arithmetic.

The visual is generated code-art, no third-party images.