Mistral launches Voxtral 3: voice agents, pricing and what to check before taking calls

What changed in the October 20 release?

Mistral describes Voxtral 3 as a speech-to-text and text-to-speech family, with an SDK connecting listening, reasoning, tool calls and spoken replies. Its announcement offers an EU-hosted API and self-deployment options. The dated model cards mark the new models as public preview.

That distinction matters when planning a service customers will depend on. Before choosing a launch date, confirm account access, the specific model version, support terms and the deployment route you intend to use. Promptwire has not independently tested the new models.

Public record: Mistral — Voxtral 3 announcement · Mistral — Small Transcribe 3 model card, dated October 20, 2026 · Mistral — Mini Transcribe 3 model card

Choose recorded transcription or live conversation first

Mistral’s audio overview separates recorded-file transcription, live streaming transcription, speech generation and streaming speech generation. The Small transcription model is described there as transcription-only at launch. This is narrower than its model card’s broader mention of translation, speech questions, summarization and function calling.

Treat those as separate integration questions. A meeting-transcript pilot can work from a completed recording. A live service desk also needs to decide when a caller has stopped speaking, manage interruptions and keep business-system actions consistent with the conversation. Test the actual endpoint behavior rather than assuming every listed feature works in every mode.

Public record: Mistral — audio overview and model selection · Mistral — Small Transcribe 3 model card, dated October 20, 2026

The published prices use different billing units

At the time of checking, the Small transcription card lists $0.004 per audio minute, Mini transcription lists $0.0035 per minute, and Mini Realtime transcription lists $0.008 per minute. The TTS 3 card lists $13 per million characters.

For 1,000 billed audio minutes, those transcription rates imply $4, $3.50 and $8 respectively. These are arithmetic examples for the individual transcription components. A full voice agent can also incur reasoning-model, speech-generation, telephony and integration costs. Do not compare a transcription-only bill with the price of a complete call-handling service.

Component rates from Mistral model cards; check current billing terms before estimating production cost.

Public record: Mistral — Small Transcribe 3 model card, dated October 20, 2026 · Mistral — Mini Transcribe 3 model card · Mistral — Mini Realtime Transcribe 3 model card · Mistral — TTS 3 model card

Check the language list for the exact model

The announcement says transcription covers more than 50 languages and presents speech generation as supporting more than 35 languages. However, the TTS 3 model card currently says nine languages. Those public descriptions do not agree. Confirm the language and dialect you need in the chosen endpoint before promising coverage to customers.

The Realtime transcription card describes configurable latency from 80 to 2,400 milliseconds, speaker labeling, vocabulary biasing and end-of-turn detection. That is a component setting, not a measured guarantee of how quickly an entire voice agent will respond. Network delay, reasoning and tool execution still sit in the path.

Public record: Mistral — Voxtral 3 announcement · Mistral — TTS 3 model card · Mistral — Mini Realtime Transcribe 3 model card

A practical test: one corrected number, one interrupted request

Build a short test set around the details that cause expensive rework. Include a caller correcting an order number, a change of date halfway through a request, two people speaking, background noise and an unfamiliar product name. If your callers mix languages, include that pattern with reviewers who understand both.

Write the expected outcome beside each recording. The agent should retain the corrected number, read back the final date and avoid claiming a booking succeeded when the connected system returned an error. Test a timed-out tool call twice to check that the workflow does not create duplicate requests.

Score the complete job: factual accuracy, successful tool outcome, interruption handling, escalation quality, elapsed time and human correction effort. Listen to failures rather than relying only on a transcript. A fluent response can conceal a misheard number.

Start with a narrow demonstration against test records. Expand only after the measured results satisfy your own acceptance criteria and the documented access and language details match the intended use. This evaluation method is Promptwire’s editorial guidance, not a reported benchmark result.