Voxtral 3 is a phone bill, not a checkpoint: Mistral sells the whole voice pipeline

Mistral published Voxtral 3 on 20 October: a family of transcription and speech models plus a voice agent SDK that strings them into one streaming pipeline. The models are the headline; the SDK is the product, priced like a telephone line rather than a licence.

The pipeline is the pitch

The launch post describes an agent that listens, reasons, calls a tool and speaks with no stage waiting on the last, aiming at under a second per turn and stopping mid-sentence when the caller cuts in. It is built on LiveKit, answers and places calls over SIP, runs in browsers over WebRTC, and plugs into existing contact-centre software. Because each stage is a separate model, the reasoning LLM, voice or language can be swapped without rewriting the agent.

The pricing makes the intent plain. The agent API costs $0.04 per minute when Mistral Small 4 does the reasoning, or $0.08 with another vendor's model. Adopting it means buying telephony minutes, not weights to host.

What the charts show, and what they skip

Transcription runs on Voxtral Small Transcribe 3: 50 languages including 24 official EU ones, with diarization, keyphrase biasing and end-of-turn detection. In Mistral's own FLEURS 45 chart it scores 4.0 to 4.1 per cent macro-average word error rate, against 4.9 for Scribe v2, 4.5 to 6.1 for Gemini 3.5 Transcribe, 5.6 to 6.3 for GPT Transcribe and 14.9 to 16.0 for Whisper large-v3. Diarization is closer: 18.4 per cent cpWER against 19.0 for Pyannote Orchestrator. The most consequential figure may be keyphrase recall on Earnings 22, 94.6 per cent with supplied vocabulary against 81.4 without, the difference between a transcript that recognises a ticker symbol and one that guesses. The Mini variant is 4B parameters at $0.0035 per minute.

Every comparison there was run by Mistral against competitors' APIs — ordinary practice, and still unaudited. The demo is a scripted transcript animation about a dermatology referral, not a recording of a call. A word error average says nothing about barge-in, unfamiliar accents or the tail latency that decides whether a caller hangs up.

Sovereignty is the sales argument

Deployment is the wedge: Mistral's EU-hosted API, or self-hosting on-premises and air-gapped, giving buyers "control over where audio is processed, how data is handled, and whether conversations are retained," as the post puts it. Hacker News readers read the missing download link as a closed release — "Closed model unfortunately" — a reversal from the open-weight Voxtral TTS of March. Another asked what the benchmarks cannot answer: "How real time is this conversation? They're showing something with zero lag between responses."

Hold onto that question: the European data argument is real, and it is sold alongside numbers produced by the seller, on a demo that was edited rather than dialled.