Mistral’s Voxtral Mini Transcribe 3 costs $0.0035 a minute and ranks 3rd on speaker labeling

Mistral AI has added Voxtral Mini Transcribe 3 to its models page. The entry describes strong transcription and speaker separation at a competitive price. Its model page, which shows a release date of October 20, 2026, lists it at $0.0035 per minute.

The model appeared in the “Other specialist models” group of the docs, next to three other new entries: Voxtral Small Transcribe 3, Voxtral Mini Transcribe Realtime 3 and Voxtral TTS 3. This article covers only the first of them, because it is the one with a detailed page. At the time of writing, Mistral had not published a blog post or changelog entry for the release.

What the listing says about the model

The API name is voxtral-mini-transcribe-3-0, version 3.0, and the model is marked Premier, Mistral’s label for its commercial models (as opposed to the Apache 2.0 open-weight ones). The model page calls it a public preview. It is served through the /v1/audio/transcriptions endpoint, has a 32k context window, and returns transcriptions with timestamps.

The page describes performance as strong and says the model is small enough for resource-constrained environments. It also says the model ranks 3rd on the FLEURS multilingual transcription benchmark and 3rd on an English diarization benchmark. Diarization means labeling who spoke when. The ranking claims are Mistral’s own, and the page does not say which models rank above it, how many were compared, or what scores were measured.

Neither the models page nor the model page gives supported languages, file size or duration limits, audio formats, or rate limits. A euro toggle appears next to the price, but no euro figure was shown. Voxtral Small Transcribe 3 is described only as the flagship transcription and diarization model with state-of-the-art accuracy, with no price on the models overview.

A small price step up from Transcribe 2

The closest comparison is Mistral’s own previous model. When the company introduced Voxtral Transcribe 2 in February, Voxtral Mini Transcribe V2 cost $0.003 per minute and Voxtral Realtime cost $0.006. It covered 13 languages and reported a word error rate of about 4% on FLEURS, according to that post. The new model’s $0.0035 is roughly 17% more per minute than V2, though the listing does not say whether it replaces V2, which still appears in the same docs as a Premier model.

For developers, the practical point is that diarization is part of the pitch for the Mini tier. In the V2 announcement, Mistral said the model adds speaker labels with start and end times, and that in overlapping speech it typically transcribes only one speaker. Whether version 3 improves on that is not stated. The 3rd-place diarization ranking suggests competitors score higher on at least some tests, so buyers who depend on meeting transcripts with several speakers should wait for named comparisons before switching.

Mistral keeps splitting audio into tiers

The listing fits a pattern of Mistral filling out product families with several sizes and price points. Last week the company published its Timestral time series models in four sizes from 5M to 209M parameters, which we covered in Mistral’s Timestral tops three forecasting benchmarks with just 209M parameters. Audio now follows the same logic: a Mini and a Small transcription tier, a realtime model and a text-to-speech model, all with a 3.0 label.

The 3.0 numbering is also worth noting. Mistral’s earlier audio releases used date-based versions such as 26.02 and 26.03, and the Voxtral TTS model appears on the same page in both a 26.03 open-licensed form (CC BY-NC 4.0) and a new Premier 3.0 form. The page does not explain how the two relate.

What the source does not establish is substantial: no release announcement, no word error rates, no language list, no licensing terms beyond Premier, and no statement on when preview ends. Because the entry was spotted on a live documentation page, details may change before an official announcement.

Who can try it now

Anyone with access to Mistral’s API can look for the voxtral-mini-transcribe-3-0 identifier on the audio transcription endpoint. The model page labels it a public preview, which usually means behavior and pricing can still change. Teams running the older Voxtral Mini Transcribe 2 do not need to act: the deprecation table on the same page lists only the original July 2025 Voxtral Mini as retiring, on May 31, 2026, with Transcribe 2 as its replacement.