Voxtral 3: Speech Models, Prices and Preview Limits

On October 20, 2026, Mistral's changelog recorded the release of the Voxtral 3 audio family, four models that are all marked Public Preview, and a managed Voice Agents pipeline on the same day. The launch page itself carries an October 22 timestamp, so you will see both dates in coverage. The changelog is the earlier, more specific record, and it is the one we use.

The useful thing to know first is that "Voxtral 3" is not one product. It is a 4B-parameter offline transcription model, a larger flagship transcription model that also handles who-said-what, a streaming transcription model with adjustable latency, and a text-to-speech model. Each has its own API identifier, its own price, and its own place in a voice pipeline. A team that reads the launch post as a single announcement tends to budget for the wrong one.

This post is about those four components and what it costs to run them. The managed pipeline that strings them together, Mistral's Voice Agents, gets a single section here because it is a different decision: whether to let Mistral run the whole turn loop for you, billed per conversation minute, rather than call the models yourself. Our Mistral Voice Agents deployment guide covers that runtime in full. If you are comparing Mistral's text-to-speech against the rest of the market, our earlier Voxtral TTS comparison covers the previous generation and the vendors it was measured against.

The Mistral changelog entry for October 20 lists the four models by identifier and describes each in a line. That line is the clearest statement of what each model is for, and it is worth more than the launch post's narrative, which moves between models without always naming them. The table below puts the identifiers next to the jobs and the list prices.

A naming point that trips people up: the text-to-speech model is called Voxtral TTS 3 on the pricing page, in the changelog and on its model card, and Voxtral Mini TTS 3 in the launch post, and its identifier is voxtral-mini-tts-3-0. They are the same model. The two offline transcription models share a description in the changelog (offline transcription and diarization with timestamps) and differ in size and accuracy, which section 02 takes up.

The pricing page also carries a second column, headed "Cached input", at one tenth of each rate: $0.00035 a minute for Mini Transcribe 3, $1.30 per million characters for TTS 3. The page does not explain how caching applies to audio or speech requests, so we print the headline rates and leave that column for Mistral to document. If your workload would re-send the same audio or the same text often, ask before you assume the discount applies.

Two models transcribe files. The launch post gives the Mini variant a size, 4B parameters, and a price point, and gives the Small variant the accuracy claims: Mistral says it leads on multilingual transcription and diarization, naming English, French and Spanish, and that it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe. The Small model card extends the list of comparators to Microsoft AI Transcribe 2, Whisper large-v3 and GPT-Transcribe. Both models claim 50 languages, including the 24 official EU languages.

The price gap between them is small: $0.0035 against $0.004 a minute, about 14% more for the flagship. That changes the usual decision. With most vendors the small model exists to save money; here the saving over a 10,000-minute month is $5. The reason to pick Mini Transcribe 3 is throughput and latency on a large backlog, not cost. If accuracy on names, accents or crosstalk matters at all, start with Small.

Both models support what Mistral calls context biasing: you pass a list of key terms or free text, and the model favours those spellings. The launch post's keyphrase-recall chart, measured on an earnings-call benchmark with up to 100 distractor terms mixed in, is the evidence Mistral offers that this works without fine-tuning. It is a chart without printed figures, so we cannot quote a number. The failure case it guards against is familiar to anyone who has transcribed a sales call: a product name rendered three different ways in one transcript, each one wrong.

Voxtral Mini Transcribe Realtime 3 is the component a live voice agent actually calls. Its model card describes a single streaming session over WebSocket with three things a live call needs: configurable latency from 80 ms to 2,400 ms, built-in end-of-turn detection, and diarization and context biasing inside the stream. The launch post adds a claim of sub-500 ms latency in more than 35 languages, which is Mistral's figure rather than an independent one.

End-of-turn detection is the feature to understand. A voice agent has to decide when the caller has stopped speaking, and a fixed silence timer gets it wrong both ways: too short and the agent interrupts a caller who paused to think; too long and every turn feels sluggish. Mistral says its detector uses acoustic and semantic signals together, so a trailing "and, um" holds the floor while a finished sentence releases it. Whether it does that well on your callers is a test, not a given.

The configurable latency is a trade-off, not a free setting. At the low end the model commits to words sooner and has less context to correct them; at the high end transcripts are steadier but the agent waits. Mistral does not publish an accuracy curve against that setting, so budget a short experiment across two or three values with your own recordings. And note what the 80 ms figure is not: it is the transcription stage's latency, not the time a caller waits for an answer. The full turn adds the language model and the speech model, which our voice agent latency reference explains how to measure end to end.

Voxtral Mini TTS 3 is the family's one speech-generation model. Its model card lists 35+ languages and 50+ dialects, streaming with about 90 ms time-to-first-audio, five output formats, 34 emotion tags and two pace tags, with both zero-shot and professional voice cloning. The price fell from $16 to $13 per million characters against the previous Voxtral TTS, which stays on the pricing page at the old rate. For a contact centre that generates tens of millions of characters a month, that is a real line-item change.

The feature Mistral leads with is contextual delivery: the model reads the surrounding conversation and chooses emotion and pacing without an explicit tag. The explicit tags remain for when the default is wrong. The launch post's example is a support agent reading back an email address, and the point of the example is that numbers and unique names are where synthetic voices fail in customer-facing use: a digit swallowed, a surname stressed on the wrong syllable. Mistral's own evaluation for this, Voice Code Bench by BeSimple AI, is 300 English prompts judged by a Gemini model, generated on October 20, 2026. Vendor-run, and narrow.

Now the limit that matters for a multilingual rollout. Preset voices, the ones you can use without supplying audio, cover 18 languages per the launch post: Arabic in two variants, Brazilian Portuguese, Catalan, Mandarin, Dutch, English in two variants, French, German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish and Vietnamese. Every other language in the 35+ claim needs a cloned voice. Zero-shot cloning takes a short sample; the professional tier Mistral calls Pro Voice needs about 30 minutes of audio and is pitched at reproducing a specific speaker's character across languages. For a brand voice in, say, Polish, you are commissioning a recording session before you ship.

One more comparison point, since it comes up: the previous generation shipped open weights and our earlier open-source Voxtral TTS guide covered self-hosting it. TTS 3 has no public weights as of October 26. If your reason for choosing Voxtral last time was running it on your own GPU, that reason does not yet carry over.

The per-minute and per-character prices only become a budget when you attach them to a workload, and the two units do not add up without assumptions. Here is a hypothetical, stated so you can change the inputs: a support line handling 10,000 minutes of calls a month, where the agent speaks for about half of each call at a conversational 900 characters a minute, and every call is also transcribed afterwards for quality review. The arithmetic is ours, on Mistral's list prices.

Two things the table shows. First, the speech models are not the expensive part of a voice agent. The language model's tokens, the telephony minutes and the engineering time to run the turn loop usually dwarf $180 a month. Second, the managed Voice Agents rate is not a like-for-like comparison with the components: its $0.04 a minute is billed per minute of conversation with Mistral Small 4 as the reasoning model, while the component total above excludes any language model. The pricing page does not itemise what the minute rate includes, so confirm that before treating it as an all-in price. Whether $400 for a run pipeline beats $180 plus your own model bill and your own code is the subject of the managed-runtime post. The point here is only that the units differ, which our guide to characters, tokens and minutes walks through in detail.

A failure case from the same arithmetic: a team that transcribes every call live and then re-transcribes the recording with the realtime model for review pays $0.008 twice. Use the offline model for the second pass. It is half the price, and it is the model Mistral positions as its flagship for accuracy.

All four model cards carry the same label: Public Preview. Mistral used the same label for Mistral Large 4 on October 6, and the pattern is that preview models are generally callable and priced but not yet declared stable. For a procurement conversation that matters in three ways: prices can move before general availability, identifiers and defaults can change, and the support commitments a production contract needs may not be in place yet. None of that is stated on the cards; it is the ordinary meaning of a preview label, and the prudent reading until Mistral says otherwise.

Weights are the second open item. Mistral's earlier Voxtral models were open-weight, and that history leads people to assume the new ones are too. As of October 26, the newest Voxtral repository under Mistral's Hugging Face organisation is an Arabic realtime variant dated October 8, with nothing for generation 3. The launch post offers self-deployment, but through a contact form, which means a commercial conversation rather than a download. If you need the model inside your own perimeter for data-residency reasons, that conversation is the path, and the hosted alternative is Mistral's European endpoint.

The third is where the models run. The launch post lists European, US and global endpoints for the hosted API. A European endpoint with a European vendor is a meaningful option for teams whose call recordings cannot leave the region, and it is the strongest practical argument for Voxtral over a US-hosted equivalent. It is not the same as self-hosting, and a data-protection review should treat the two differently.

Mistral released Voice Agents alongside the models: a managed pipeline that captures audio, transcribes it with the realtime model, generates an answer with a text model, speaks it with TTS 3 and plays it back, with interruption handling and tool calls. The overview page is clear that each stage is a separate model in a chain, and that each one adds to the time the caller waits. It chains transcription, a text model and speech generation, rather than using a single speech-to-speech model, and that distinction matters when you compare it with the full-duplex models in our full-duplex voice model comparison.

Two facts from the overview page decide whether it is even an option for you. Telephony over SIP is available to enterprise customers at launch, through an account team, so a self-serve developer gets browser and custom-audio interfaces only. And on a phone line no code of yours runs inside the call, so client-side tools fail and anything a phone agent needs must be a server-side tool. The pricing page bills the pipeline per minute of conversation, $0.04 with Mistral Small 4 and $0.08 with larger models. Our separate post on Mistral Voice Agents takes the runtime decision from there: deployment surfaces, tool placement, interruption behaviour and what the minute price does and does not include.

Because every figure in the launch is Mistral's, the adoption decision rests on a short evaluation of your own. It does not need to be large. Fifty of your own recordings, chosen to include the accents, crosstalk and product names that have caused complaints before, will tell you more than any leaderboard. Our GPT Transcribe comparison describes the independent leaderboard figures that exist for the previous generation of models, if you want an external baseline to sit alongside.

The failure case to design against is the one voice vendors rarely show: a transcript that is 97% accurate by word error rate and still unusable, because the 3% it got wrong were the account numbers, the surnames and the medication names. Word error rate weights every token the same. Your business does not. Build the test set so the terms that matter are over-represented, and score those separately.

If those four pass, the remaining questions are commercial rather than technical, and section 06 lists them. If you would rather have a second pair of hands on the evaluation, our AI transformation team runs exactly this kind of before-you-switch test for voice and document pipelines.