Mistral's Voxtral 3 prices a full voice agent at $0.04 a minute

Mistral released Voxtral 3 on 20 October 2026: three speech-to-text models, one text-to-speech model and a voice agent SDK that joins them to a reasoning model. The company's post says the family is available now on the Mistral API through European, US and global endpoints. The number that matters for a service desk is the bundled one: Mistral lists a voice agent API at $0.04 per minute with Mistral Small 4 reasoning, and $0.08 per minute with a larger model.

We logged the release when it landed and did not write it up. This is the fuller account, with what Mistral's own charts show against it.

What it costs

• Voxtral Mini Transcribe 3 (batch): $0.0035 per minute. Mistral says the model has 4B parameters.
• Voxtral Small Transcribe 3 (batch, the flagship for transcription and speaker labels): $0.004 per minute.
• Voxtral Mini Transcribe Realtime 3: $0.008 per minute.
• Voxtral Mini TTS 3: $13 per 1M characters.
• Voice agent API: $0.04 per minute with Mistral Small 4, $0.08 per minute with a larger model.

On my arithmetic, a six-minute call through the agent API is $0.24 or $0.48. The post does not say whether that covers phone carriage, and it gives no price for running the stack on your own hardware.

How the agent is put together

The SDK is built on LiveKit, the post says. The realtime model listens and predicts when the caller has finished, a general model such as Mistral Small reasons and calls tools, and the speech model replies. Mistral says each turn completes in under one second and that the agent stops talking when the caller cuts in. It takes and places calls over SIP and runs in browsers over WebRTC.

Each stage is a separate model, so, according to Mistral, a team can change the language model, the voice or the language without rewriting the agent. Transcription covers 50 languages, including 24 official EU languages; the speech model covers more than 35, with preset voices in 18. Both transcription models accept key terms and free text to improve recall of domain vocabulary without fine-tuning.

What Mistral's own charts show

• Speaker labelling: Voxtral Small Transcribe 3 scores 18.4% error (cpWER, lower is better) across five public benchmarks, against 20.3% for Microsoft AI Transcribe 2 and 38.4% for Gemini 3.5 Transcribe. Mistral ran the test.
• Key terms: on the Earnings 22 benchmark, recall of key phrases rises from 81.4% to 94.6% when Voxtral Small is given the terms in advance.
• Realtime: the chart does not put Mistral first. Gemini 3.5 Realtime Transcribe scores 4.2% word error rate over 35 languages. Voxtral's realtime model scores 5.7% at a 2,400 ms delay, 6.0% at 960 ms, 6.4% at 480 ms and 7.4% at 240 ms.
• Speech: the price-against-quality charts are labelled internal evals, or use a benchmark from BeSimple AI, and all are judged by a Gemini model.

The realtime row is the trade to plan for. The post advertises transcription at under 500 ms, and on Mistral's chart the 480 ms setting gives up 0.7 points of word error rate against the slowest setting, and 240 ms gives up 1.7.

What the post leaves out, and the pushback

The post names no customer and gives no containment rate, error rate on live calls or measured end-to-end latency. It does not state a licence or publish weights. Its summary describes an EU-hosted API while the availability section lists European, US and global endpoints, so confirm the region before sending audio.

On Hacker News, where the launch thread (item 50168396; news.ycombinator.com is not an allow-listed article source, so it is not linked here) drew 770 points and 202 comments, the most repeated objection was that the models are closed. One commenter wrote that nothing for Voxtral 3 is on Hugging Face and that self-deployment goes through sales; they said the last open speech model from Mistral was Voxtral Mini 4B Realtime in February, under Apache 2.0. Another raised the security question: a caller can try to talk an agent wired to internal systems into an action, such as authorising a payment, so tool permissions need the same limits as any other agent.