Voxtral 3: Mistral Puts Four Speech Models and a Voice-Agent API on the Same Meter

The interesting number in Mistral's October 20 release is not a word error rate. It is $0.04 per minute. That is what the company charges for a Voxtral 3 voice agent — an end-to-end pipeline that hears a caller, reasons about what they said with Mistral Small 4, calls tools, and talks back. Add up the parts Mistral sells separately and the same minute costs more: Voxtral Mini Transcribe Realtime 3 at $0.008, plus Voxtral Mini T​TS 3 at roughly $0.02 for a normal speaking rate, plus the reasoning model on top. The bundled minute is the product. Everything else — Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, Voxtral Mini Transcribe Realtime 3, Voxtral Speech 3 — is sold as a component you can buy à la carte, and as a way of showing that the components are good enough to be worth bundling. Voice-agent pipelines fail in ways rate cards never mention, and the recorded incidents are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What is actually new, model by model

This is a family refresh, not one model with a version bump. Each piece has a distinct job and a distinct price, and the split matters more than the shared name.

• Voxtral Small Transcribe 3 — the accuracy tier. $0.004 per minute, 32k context, and Mistral's flagship claim for transcription with diarization. It also does translation, speech Q&A, summarization and function calling, which makes it the one member of the family that is closer to an audio-language model than a transcriber.

• Voxtral Mini Transcribe 3 — 4B parameters, $0.0035 per minute, 32k context. Mistral gives this one's size and withholds the Small model's. That asymmetry is worth noticing: the mini model is the one being positioned for volume and edge deployment, so its footprint is a selling point; the flagship's is not disclosed.

• Voxtral Mini Transcribe Realtime 3 — streaming over WebSocket at $0.008 per minute, with latency configurable from 80 ms to 2,400 ms and end-of-turn detection that combines acoustic and semantic cues rather than waiting for silence. The docs say "sub-500 ms"; the configurable floor is far below that, and the accuracy cost of going low is significant — see the benchmark section.

• Voxtral Speech 3 (sold in the docs as Voxtral Mini T​TS 3) — $13 per million characters, around 90 ms to first audio, 34 emotion tags and two pace tags, nine languages, zero-shot and "professional" voice cloning. Mistral's launch page puts cloning at 30+ minutes of reference audio.

Underneath sits the voice-agent SDK, built on LiveKit and streaming ASR, LLM and T​TS as one pipeline. It takes SIP calls, WebRTC sessions and existing telephony or IVR trunks, supports barge-in, and lets you swap the reasoning model, the voice or the language without rebuilding the agent. Mistral quotes turn latency under one second.

The benchmark table, read carefully

Mistral published its own numbers, and they are good numbers. On FLEURS 45 macro-average word error rate, Voxtral Small Transcribe 3 lands at 4.1% with forced language and 4.0% with inferred language. Voxtral Mini Transcribe 3 is at 4.6% / 4.5%. For comparison on the same vendor table: G​emini 3.5 Transcribe at 6.1% / 4.5%, Scribe v2 at 4.9%, G​PT Transcribe at 6.3% / 5.6%, and Whisper large-v3 at 16.0% / 14.9%.

Two things about that table deserve a second look. First, Microsoft's MAI-Transcribe-2 sits at 4.1% / 4.0% — identical to Voxtral Small Transcribe 3, down to the decimal. A vendor table in which the flagship ties its closest competitor exactly is either an honest tie or a rounded one; Mistral does not say which. Second, the Whisper figure is not the number anyone else quotes for Whisper. Whisper large-v3 scores 7.44 mean WER on the Open ASR Leaderboard, and the Turbo variant 4.6% on Artificial Analysis's independent board. Sixteen percent is what Whisper looks like on hard multilingual FLEURS audio specifically, not what it looks like in production English. Using the harshest available slice for a competitor is a normal marketing move, but it is a marketing move.

The independent board tells a tighter story. Artificial Analysis has Voxtral Small Transcribe 3 at 2.1% AA-WER and Voxtral Mini Transcribe 3 at 2.2% — first and second on the non-streaming leaderboard as of this writing, with MAI-Transcribe-2 at 2.0% in the slot above them depending on when the board is read. Those are close enough that the ordering will move on reruns. Anyone quoting a winner from a 0.1-point gap is quoting noise.

Where the realtime tier gives accuracy back

The streaming numbers are the most useful thing Mistral published, because they are the least flattering. On FLEURS 35, the offline Voxtral Mini Transcribe 3 baseline is 4.0% WER. The realtime model degrades monotonically as you ask it to answer sooner:

• 2,400 ms latency — 5.7% WER

• 960 ms — 6.0%

• 480 ms — 6.4%

• 240 ms — 7.4%

Against that, Mistral lists G​emini 3.5 Realtime Transcribe at 4.2%, Microsoft's streaming tier at 5.8%, G​PT Live Transcribe at 8.4% and Scribe v2 Realtime at 8.7%. So the honest reading is: at a comfortable 2.4-second lag the Voxtral realtime tier is competitive with Microsoft and beats the O​penAI and ElevenLabs streaming models, but it does not beat G​oogle's, and at 240 ms it is worse than everything except the two trailing models. The sub-500 ms headline is real; it just costs you roughly three points of accuracy relative to running the same audio offline.

Diarization is where the family looks strongest and where the vendor table is doing the most work. On concatenated minimum-permutation WER, Mistral reports 18.4% for Voxtral Small Transcribe 3, 18.6% for the Mini, 19.0% for Pyannote Orchestrator Precision 3, 20.3% for MAI-Transcribe-2 and 38.4% for G​emini 3.5 Transcribe. A 20-point gap over G​emini on speaker attribution is the kind of number that decides a product — meeting transcription, call QA — but it is one vendor's harness against another vendor's model, and no third party has rerun it.

What the pricing buys, in hours

Per-minute billing is easy to compare and easy to get wrong when the alternative is per-token. Converted to a thousand audio hours, which is roughly a mid-size contact centre's month:

• Voxtral Mini Transcribe 3 — $210 per 1,000 hours at $0.0035/min

• Voxtral Small Transcribe 3 — $240 per 1,000 hours at $0.004/min

• Voxtral Mini Transcribe Realtime 3 — $480 per 1,000 hours at $0.008/min

• Voice agent API with Mistral Small 4 reasoning — $2,400 per 1,000 hours of conversation at $0.04/min

The agent tier is five to eleven times the price of the transcription tier, which is the correct order of magnitude once you remember it includes the reasoning model and the speech synthesis. It is also where the margin is, and Mistral's positioning — identity verification, bill checking, appointment booking — is chosen for calls where one completed task is worth more than a minute of compute.

One structural caution on the whole price list: every model here is in public preview, not GA, and preview pricing in speech has historically moved. There is no published commitment that $0.0035 stays $0.0035.

What Mistral did not ship

No weights. The launch page says Voxtral 3 is available through Mistral's API with EU, US and global endpoints, and "for self-deployment on your own infrastructure" — but it publishes no licence, no parameter counts for three of the four models, and no Hugging Face repository. For context, Mistral's earlier speech releases were downloadable: Voxtral Small 24B 2507 is Apache 2.0 on Hugging Face, Voxtral 4B T​TS 2603 is CC-BY-NC-4.0, and Voxtral Mini 4B Realtime Arabic went up on October 8 with no licence declared at all. This release is the first Voxtral generation with no downloadable member. Self-deployment here reads as an enterprise contract, not a download.

Also absent: any maximum audio length, any context figure for the T​TS or agent tiers, and any retention or data-residency detail beyond the endpoint regions. If you are evaluating this for regulated call recording, those three gaps are the ones to close in a procurement conversation rather than from the launch post.

How to actually try it

Voxtral 3 is not in OrcaRouter's catalogue today, so the route to evaluating it is Mistral's own API and Studio playgrounds, which cover transcription, speech generation, voice cloning and the agent pipeline. Where a router does help is the second half of the decision: you almost never want one speech stack, and putting the reasoning model that handles your agent's tool calls behind a single endpoint means you can compare Mistral Small 4 against a dozen other models on the same transcript without rebuilding the pipeline. OrcaRouter passes provider list prices through at 0% markup and fails over between providers automatically, which matters most in exactly this situation — a preview-tier model with no uptime history in front of it.

The practical evaluation order, if you are doing this this week: run your own worst audio through Voxtral Small Transcribe 3 and MAI-Transcribe-2 on the same 200 files, because the two vendors print the same FLEURS score and only your audio can break the tie. Then run the realtime tier at 960 ms and at 2,400 ms and count how often the end-of-turn detector cuts a caller off mid-sentence — that failure mode costs more in a live agent than any WER difference in this family. Then price the agent minute against your containment rate, not against your transcription bill.

What to watch

Three things will decide whether this release ages well. Independent reruns of the diarization claim, which is the only number in the launch that would genuinely change a buying decision. Whether the weights appear, because a downloadable Voxtral would immediately become the default open option for European-language diarization and would undercut the entire pricing argument above. And whether the sub-second agent latency holds on real telephony rather than in a demo of a French referral letter — barge-in and end-of-turn detection are where voice agents usually fail, and neither survives a vendor blog post as evidence.