Mistral's Voxtral 3 Lands a #3 Transcribe Model and a $13 Voice

The number to lead with is 2.1%. That is the word error rate Artificial Analysis measured for Voxtral Small Transcribe 3, which puts Mistral's new recogniser third on the independent speech-to-text board as of October 21, 2026 — behind only A​libaba's Fun-Realtime-ASR-preview at 1.7% and Microsoft's MAI-Transcribe-2 at 2.0%, and ahead of ElevenLabs Scribe v2 at 2.2%. The same October 20 release also shipped Voxtral 3 T​TS, a synthesis model priced at $13 per million characters that lands thirteenth on the Provider Voice arena and fourteenth on the Controlled Voice arena. One of those results is a genuine move up the table. The other is a competent mid-pack entry with an aggressive price tag, and the gap between how the two were announced and how they actually scored is the most useful thing in the launch. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Everything below separates three kinds of claim: numbers from Artificial Analysis's blind-listening and transcription boards (independent), numbers from Mistral's own launch page (vendor-reported), and arithmetic done here from published prices. They are not interchangeable, and the launch page presents them in a way that makes them easy to conflate.

The transcription result is the one that actually moves Mistral up the board

Voxtral Small Transcribe 3 is the flagship of the four-model family, and it is the only part of the release with an independent score that beats the incumbents outright. At 2.1% WER it is 0.1 points behind Microsoft's MAI-Transcribe-2 and 0.1 ahead of Scribe v2 — margins well inside the noise of a single evaluation set, but enough to put Mistral in the conversation at the top of the table rather than in the chasing pack.

The speed figure is the more surprising half. Artificial Analysis measures a median speed factor of 66.6 for Voxtral Small Transcribe 3, against 20.3 for the top-ranked Fun-Realtime-ASR-preview and 308.8 for MAI-Transcribe-2. In plain terms, Mistral's model transcribes roughly 67 minutes of audio per minute of compute — slower than Microsoft's by a factor of four, faster than A​libaba's by a factor of three. Cheap, accurate and moderately slow is a perfectly good shape for batch transcription; it is a poor shape for anything interactive, which is why Mistral sells a separate realtime model.

The family, as shipped on October 20:

• Voxtral Small Transcribe 3 — 50+ languages, diarization, context biasing, $0.004 per minute ($4.00 per 1,000 minutes). The accuracy pick.

• Voxtral Mini Transcribe 3 — 4B parameters, 50+ languages, $0.0035 per minute. Scores 2.2% WER on the same board — statistically level with the Small model at 12.5% less money, which makes the Small tier's premium hard to justify on accuracy alone.

• Voxtral Mini Transcribe Realtime 3 — streaming, sub-500ms latency claimed, $0.008 per minute. Mistral's own page says both "35+ languages" and "40+ languages" for this model in two different places; nobody outside has tested which is right.

• Voxtral 3 T​TS — batch and realtime synthesis at $13 per million characters, 35+ languages and 50+ dialects. The model Mistral's pricing table calls "Voxtral Speech 3" and its documentation lists as voxtral-mini-tts-3-0.

On Mistral's own FLEURS 45 evaluation, Voxtral Small Transcribe 3 reports 4.1% WER with forced language and 4.0% with inferred language. That is a different benchmark from the one Artificial Analysis runs, which is why the vendor page shows 4% and the independent board shows 2.1%. Neither number is wrong; quoting them in the same sentence without saying which is which is how launch coverage goes wrong. Mistral also claims the lowest diarization error in its comparison set at 18.4% cpWER and 94.6% keyphrase recall with context biasing, up from 81.4% without — vendor-reported, unreproduced.

The voice model is mid-pack, and the price is doing the arguing

Mistral's social post frames the T​TS launch as entering the top twelve of both text-to-speech arenas. The live boards on October 21 tell a slightly softer story: Voxtral 3 T​TS sits at #13 on Provider Voice with 1,217 Elo (95% interval 1,202–1,232, a tied band spanning ranks 11 through 17) and #14 on Controlled Voice with 1,066 Elo (a tied band spanning 10 through 19). It is, in other words, in the top fifteen on both — and the confidence intervals overlap with five other models, so the exact rank is not a stable fact. This is the model's second day on the board and its sample count, 1,619 comparisons, is well below the 3,000+ that the leaders have accumulated.

What is stable is the neighbourhood. At 1,217 Elo, Voxtral 3 T​TS is level with Realtime T​TS-2 Flash and StepAudio 2.5 T​TS, ahead of ElevenLabs Eleven v3 at 1,182, and behind Cartesia Sonic 3.6 at 1,289, Q​wen-Audio-3.1-T​TS-Plus at 1,299 and ElevenLabs Eleven v4 at 1,328.

Two things about that placement are worth saying plainly. First, Voxtral 3 T​TS is a clear generation-over-generation improvement for Mistral: the March 2026 Voxtral T​TS scored 1,090 on the same board, so this is a 127-point jump. Second, the jump did not close the gap to the frontier. Mistral's previous model was a long way back; the new one is in the middle of the pack.

What $13 a million characters actually buys

Price is where this release gets interesting, and where a straight reading of the press release will mislead you. At $13 per million characters, Voxtral 3 T​TS costs about $0.70 per hour of generated speech, assuming roughly 900 characters per spoken minute. Cartesia Sonic 3.6 runs $2.65 an hour at $49, ElevenLabs Eleven v4 runs $4.32 at $80, and Q​wen-Audio-3.0-T​TS-Plus runs $1.04 at $19.31. Against the premium tier the saving is real and large.

Against the value tier it is not a knockout. Simba 3.2 sits at #10 with 1,245 Elo — 28 points above Voxtral 3 T​TS — at $6.6 per million characters, half the price. Realtime T​TS-2 Flash scores the same 1,217 Elo at $10.4. Gemini 3.8 Flash-Lite T​TS scores 1,250 at $11. On a crude Elo-per-dollar ratio, which treats Elo as linear when it is not and should be read as a rough sort rather than a measurement, Voxtral 3 T​TS lands at about 94 points per dollar against 189 for Simba 3.2 and 118 for Realtime T​TS-2 Flash.

The honest summary is that Mistral priced the model to be defensible rather than disruptive. It undercuts every model that beats it by more than 50 Elo, and it loses on value to at least three models that beat it on quality. That is a normal launch, not a repricing event.

Where routing matters here is in keeping that comparison live rather than in a spreadsheet. OrcaRouter runs one API for 200+ models and passes provider list price through at 0% markup, so when a vendor like Mistral moves a rate the number your code sees moves with it, and swapping the synthesis endpoint behind a voice pipeline is a config change rather than a second contract. Whether Voxtral 3 T​TS belongs in that pipeline is a separate question from whether it is the best model on the board — and on the board, today, it is not.

The part of the release that is not a model

The most consequential thing Mistral shipped on October 20 may not be either scored model. The family comes with a voice-agent SDK built on LiveKit, supporting SIP telephony, WebRTC, IVR and contact-centre integrations, with a claimed full agent turn under one second and barge-in handling. Mistral prices the whole agent loop — speech in, reasoning, speech out — at $0.04 per minute when the reasoning model is Mistral Small 4, or $0.08 per minute with a different reasoning model.

That per-minute bundle is a different unit from a per-character T​TS rate, and comparing the two directly is a common mistake. A voice agent that talks for 30 seconds of a one-minute call consumes roughly 4,500 characters of synthesis, which at $13 per million is under six cents — so the $0.04 bundle is competitive with assembling the parts yourself, provided you wanted Mistral's reasoning model anyway. The bundle is also where the failover argument earns its keep in production: a voice agent with a dead synthesis endpoint is a silent phone line, and the reasoning leg of that pipeline is exactly the kind of thing a router with automatic failover is for.

What Mistral did not ship is open weights. The Voxtral 3 page describes API access, Studio playgrounds and self-deployment on private infrastructure, but never states that the weights are downloadable. The March Voxtral T​TS shipped on Hugging Face under CC BY-NC 4.0; nothing in this release repeats that. If self-hosting is a requirement, this generation does not currently answer it.

Three things nobody has measured yet

The absence of a published latency figure for Voxtral 3 T​TS is the sharpest gap in the launch material. Mistral gives sub-500ms for realtime transcription and under one second for a whole agent turn, but no time-to-first-audio for synthesis, which is the number that decides whether a voice agent feels conversational. Artificial Analysis reports a median generation time of 7.2 seconds and a median throughput of 112 characters per second, both of which describe batch rendering rather than streaming behaviour. A 112 characters-per-second median is respectable — faster than ElevenLabs Eleven v3's 56 and slower than Sonic 3.6's 145 — but it says nothing about the first 200 milliseconds.

Voice cloning is the second gap. Mistral's "Pro Voice" adaptation reproduces a speaker's character and emotional range from 30 minutes of audio. That is a studio-grade process, not the ten-second instant clone Cartesia advertises or the thirty-second clone ElevenLabs offers, and it positions Voxtral 3 T​TS for brand voices rather than for user-generated character work.

The third is endurance. Mistral's own materials do not state a maximum generation length for the T​TS endpoint, and the arena samples are short clips. Long-form narration is where a mid-pack Elo model can still be the right answer if it holds coherence across ten minutes, and nobody has published that test.

The fair reading of October 20 is that Mistral shipped a real transcription competitor and a T​TS model that is good enough to be on the shortlist and cheap enough to be tried. The transcription board is where the company earned a headline. The synthesis board is where it bought one.