Mistral dropped its biggest voice release yet on October 20. Voxtral 3 is a full stack for building real-time voice agents: transcription, speech synthesis and an SDK that stitches listening, reasoning, tool calls and speaking into one streaming pipeline. No gluing separate APIs together after the fact.

What's in the box

On the listening side you get three transcription tiers. Voxtral Mini Transcribe 3 handles batch jobs, Voxtral Small Transcribe 3 adds speaker diarization, end-of-turn detection and key-term biasing for jargon, and Voxtral Mini Transcribe Realtime 3 targets live agents with what Mistral says is sub-500ms latency. Coverage runs to 50-plus languages, including all 24 official EU languages.

Voxtral Speech 3 handles the talking: 35-plus languages, 50-plus dialects, and it adjusts emotion, pacing and emphasis to the conversation instead of reading flat. A Pro Voice feature clones a voice from 30-plus minutes of reference audio. The SDK itself is built on LiveKit, supports SIP phone connections and WebRTC, and lets agents call your own tools mid-conversation. Booking an appointment or pulling up an invoice while the caller is still on the line is the use case Mistral keeps coming back to.

The pricing

All approximate, as always. Batch transcription starts at $0.0035 per minute, Small at $0.004, and realtime at $0.008. Speech generation costs $13 per million characters. The full agent API, which bundles the whole loop, runs $0.04 per minute with Mistral Small 4 doing the reasoning and $0.08 with a larger reasoning model.

For context, that's genuinely aggressive. Mistral positions Speech 3 as cheaper than OpenAI's TTS-1, Deepgram and ElevenLabs Flash. If you're running a support line or a voice product at volume, the math deserves a spreadsheet.

The part we'd like verified

Then there are the benchmarks. Mistral claims 4.1% word error rate on the multilingual FLEURS test versus 16% for Whisper large-v3, plus wins on diarization. Those are Mistral's own numbers on Mistral's own launch post, with no per-language breakouts and no customer names attached. Vendor benchmarks are a starting point, not a verdict. We'd like to see independent tests before crowning anything.

Why this one matters

Two things stand out. First, deployment: you can run the stack on your own infrastructure or through Mistral's EU-hosted API, with control over whether audio is retained. For European businesses sweating data residency, that's a real differentiator. Second, timing. Mistral shipped its Timestral time-series model the day before, so this lab is moving fast.

Our take? Voice AI is quietly becoming one of the most competitive corners of the market. OpenAI, ElevenLabs, Deepgram and now Mistral are all fighting over the same developers. For you, the builder, that fight means lower prices and better tools. We'll be watching whether the sub-500ms latency claim holds up in the real world.