Voxtral 3 vs Gemini 3.5 Transcribe: Mistral Says It Wins the Words — Google Wins the Rest

On the only metric both vendors agree to print, these two are a hair apart. Voxtral Small Transcribe 3 — Mistral's flagship transcriber from the Voxtral 3 family released on October 20, 2026 — sits at 2.1% word error rate on Artificial Analysis's non-streaming board. G​emini 3.5 Transcribe, which G​oogle has run as a generally available model since 26 August 2026, sits at 2.6% on the same board. Five-tenths of a point is roughly five extra errors per thousand words. That is not the reason to pick one. The reasons are everything above the words: G​emini does diarization for eight speakers, takes a thousand custom vocabulary terms, cleans up disfluencies and resolves self-corrections in a single pass, and spans 85+ locales. Voxtral does diarization measurably better by its own table, covers 50+ languages, and undercuts G​oogle's per-minute rate. This page is about which of those differences actually reaches your pipeline. Word error rates do not record how either vendor's transcription API fails under load, and those incidents are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What each one is selling

Mistral's framing for Voxtral Small Transcribe 3 is accuracy plus speaker attribution. It is the "flagship transcription and diarization model" in the family, it does timestamps and diarization natively, and it also does translation, speech Q&A, summarization and function calling — which is a quiet way of saying it is an audio-language model that transcribes rather than a transcriber. At $0.004 per minute with 32k context, it is priced as a direct substitute for the managed transcription APIs.

G​oogle's framing for G​emini 3.5 Transcribe is what happens to the transcript after recognition. The default `verbatim` mode keeps fillers, false starts and repetitions; `smart` mode strips them and applies structured formatting; a custom vocabulary biases toward your jargon; word-level timestamps come back as annotations; and speaker labels arrive as `spk_1`, `spk_2` and so on. It runs through the Interactions API for files and a separate `gemini-3.5-transcribe-live` SKU for streaming, and it carries a free tier that Mistral does not offer.

Put simply: Mistral is selling you a better ear, G​oogle is selling you a better editor.

Dimension by dimension

• Independent accuracy — Voxtral Small Transcribe 3 at 2.1% AA-WER vs G​emini 3.5 Transcribe at 2.6%, both on Artificial Analysis's non-streaming leaderboard. Mistral's own FLEURS table prints a wider gap: 4.1% forced-language for Voxtral vs 6.1% for G​emini, but only 4.0% vs 4.5% with inferred language. The independent board agrees with the inferred-language reading, so treat the 6.1% figure as the loose end of Mistral's table rather than the headline.

• Price — $0.004/min vs $0.005/min, both blended list rates. That is $240 vs $300 per thousand audio hours. Real, but not decisive on its own; a 20% delta gets erased by one re-run of a bad batch.

• Speaker attribution — Voxtral Small Transcribe 3 at 18.4% concatenated min-permutation WER on Mistral's harness, against G​emini 3.5 Transcribe at 38.4% on the same vendor table. This is the largest gap in the launch post and the one most likely to matter. Caveat it honestly: both numbers are Mistral's, run on Mistral's harness, and G​oogle has published no equivalent diarization score to check against. G​oogle's documentation does disclose a hard ceiling that is not a benchmark — eight speakers, with attribution "experimental" for three or more — which for meeting transcription is a product limit, not a quality question.

• Language coverage — Voxtral: 50+ languages, including all 24 official EU languages and a set of low-resource Asian and Indic languages. G​emini: 85+ locales with BCP-47 hints and intra-sentence code-switching. G​oogle's list is longer and, more importantly, named — you can check whether your locale is on it. Mistral's "50+" is a count without a list.

• Audio limits — G​emini caps unary audio at one hour, or thirty minutes when you ask for diarization or word timestamps. Mistral publishes no maximum audio length for Voxtral Small Transcribe 3 at all. That silence is not a feature. If you process long files, G​emini is the only one of the two you can size a job queue against without asking sales.

• Vocabulary control — G​emini accepts up to 1,000 custom terms (best results up to about 100) and explicitly warns that custom vocabulary cannot be combined with diarization or word timestamps. Voxtral's context biasing takes key terms and free-form text, and Mistral reports keyphrase recall of 94.6% with prompting vs 81.4% without. Note the shape of that claim: it is recall-with-bias, not errors-with-bias, and it is vendor-reported.

• Self-correction handling — G​emini's `smart` mode resolves self-corrections and normalises spoken numbers and currency. Mistral offers no equivalent documented mode. If your audio is interviews or call-centre speech full of "no wait, I meant", this is the single feature most likely to decide the choice.

• Streaming — G​emini has a live SKU at roughly $0.009/min blended. Voxtral's streaming tier is a different model, Voxtral Mini Transcribe Realtime 3, at $0.008/min with latency configurable from 80 to 2,400 ms. Neither vendor's streaming model is the accuracy flagship, so comparing streaming accuracy to the batch numbers above is a category error.

• Maturity — G​emini 3.5 Transcribe has been GA since 26 August 2026 and has an independent WER score on a public board. Voxtral Small Transcribe 3 is in public preview as of 20 October 2026, with no uptime history, no GA date and no published commitment that its preview price holds.

The diarization number is the one to test

If there is a single sentence in this comparison that should send you to your own audio, it is this: Mistral claims 18.4% diarization error against G​emini's 38.4%, and G​oogle publishes no counter-number. That is a suspiciously large gap for a difference of accuracy class — 2.1% vs 2.6% on plain transcription does not predict a 2× gap on speaker attribution. Two explanations are available and only your data separates them. Either G​emini's diarization is genuinely weaker, which would fit its documented eight-speaker ceiling and "experimental" note for three or more speakers, or Mistral's harness flatters its own model in a way a neutral rerun would not reproduce.

Practical test: take forty minutes of the worst multi-speaker audio you own — crosstalk, a speaker on a phone line, at least one place where two people talk over each other — run both, and count speaker switches rather than word errors. That takes an afternoon and settles a claim neither vendor can settle for you.

What the transcripts actually cost at scale

Work the arithmetic on a support organisation recording 4,000 hours a month. At $0.004/min, Voxtral Small Transcribe 3 comes to $960. At $0.005/min, G​emini 3.5 Transcribe comes to $1,200. The $240 monthly gap is real and is roughly the cost of one engineer-hour per working day, which is the correct way to weigh it: meaningful against a self-hosted alternative whose ops cost you nothing, irrelevant if G​emini's `smart` mode removes a post-processing step you would otherwise pay for.

That post-processing step is where the price comparison usually breaks. A raw transcript with fillers and unresolved self-corrections needs a cleanup pass — either a regex layer you maintain or an LLM call per file. A cleaned, punctuated, speaker-labelled transcript from G​emini does not. If your pipeline currently spends a language model call per transcript to fix formatting, put that call in the comparison before you conclude Mistral is 20% cheaper.

For teams that want both — Mistral's diarization on the calls where speakers matter, G​emini's cleanup on the bulk archive — the constraint is usually keys and vendors rather than money. OrcaRouter's model catalogue does not currently include either Voxtral 3 or G​emini 3.5 Transcribe, so this particular pairing still means two contracts. Where we do help is the layer above: the reasoning model that summarises or routes those transcripts can come from a single endpoint with automatic failover, so the speech vendor choice does not also lock your downstream model choice.

Verdict

Choose Voxtral Small Transcribe 3 if your problem is speakers and your languages are European — the diarization claim, the 24 EU languages, and the 20% lower rate all point the same way, and 2.1% WER on an independent board means you are not trading accuracy for it. Choose G​emini 3.5 Transcribe if your problem is the transcript itself: self-corrections, formatting, a large named locale list, and a published audio-length limit you can plan capacity against. The one thing I would not do is pick on the 0.5-point WER gap. It is smaller than the run-to-run variance on any board, and both vendors print the number that suits them from the same family of evals.

The open question is what happens when Voxtral Small Transcribe 3 leaves preview. If Mistral holds $0.004 and an independent diarization rerun confirms the 18-point gap, this stops being a close call for anyone doing meeting or call analytics. If the price moves at GA, the case narrows back to language coverage and speaker attribution alone.