Mistral Voxtral 3: Transcription From $0.0035 a Minute and a Full Voice Agent API at $0.04 a Minute

Until now, a team building a phone or voice agent on Mistral had to bring its own speech layers. Voxtral 3 changes that. Mistral now offers transcription, real-time transcription, speech synthesis and an agent SDK under one bill, and it priced the full loop per minute rather than per token.

The price list

One small inconsistency: Mistral's page gives the real-time model's language count as 35+ in one place and 40+ in another. Test your target languages before you commit.

The agent SDK

• Built on LiveKit, the open-source real-time audio framework many voice agent teams already use.
• Supports SIP for phone lines and WebRTC for browser and app calls.
• Mistral says a full turn, from the caller finishing a sentence to the agent starting its reply, completes in under one second.
• End-of-turn detection in the transcription models is what lets the agent know when to speak without cutting people off.

The benchmarks, as Mistral reports them

Mistral also reports an 18.4% concatenated minimum-permutation word error rate for diarization on Small Transcribe 3, and keyphrase recall of 94.6% when you pass it context terms against 81.4% without. That second number is the practical one: if your calls are full of product names or medical terms, feeding them in as context matters more than the model choice.

What the coverage left out

• Weights: earlier Voxtral releases shipped open weights. The Voxtral 3 announcement mentions self-deployment but does not state an open-weights licence for each model, so confirm before planning an on-premises build.
• Benchmarks: every number above comes from Mistral. Microsoft ties Small Transcribe 3 on the headline chart, so the accuracy lead is not as clear as some headlines suggest.
• Full price table: most write-ups quoted one or two prices. The agent API at $0.04 per minute is the figure to compare against a stitched-together stack of separate speech, model and voice vendors.

What this means for teams building voice agents

• If you run high call volumes, $0.04 per minute for the whole loop is a simple number to model against your current per-minute cost across three vendors.
• If you already use LiveKit, trying Voxtral is mostly a provider swap rather than a rebuild.
• If data location matters, the EU endpoint is the main reason to look at Mistral over US-only providers.
• Run your own word error test on real recordings. Vendor charts on FLEURS rarely match noisy phone audio.