Mistral Voxtral 3 Review (2026): Voice Agents at $0.04/min

Three transcription models, one text-to-speech model and a voice agent SDK, released together at prices that undercut the field — with one accuracy number Mistral publishes against itself.

By AI Tool Scout · Published October 22, 2026 · 19 min read

What Is Mistral Voxtral 3?

Mistral Voxtral 3 is the October 20, 2026 release that turns Mistral's speech line into a complete stack: three transcription models, one text-to-speech model and a voice agent SDK, all served from the Mistral API on EU, US and global endpoints or self-hosted on your own infrastructure. The transcription models are Voxtral Mini Transcribe 3 (4B parameters, batch, $0.0035 per minute), Voxtral Small Transcribe 3 ($0.0040 per minute, positioned as the strongest on multilingual transcription and speaker attribution) and Voxtral Mini Transcribe Realtime 3 ($0.008 per minute, streaming, 35+ languages, under 500 ms). The synthesizer is Voxtral Speech 3, also called Voxtral Mini TTS 3 in the launch post, at $13 per million characters in batch and realtime alike.

The reason this is a bigger announcement than four model cards is the SDK. Until now, standing up a voice agent meant three vendor contracts, three billing units and three failure modes: a transcription API billed per minute, a language model billed per token, and a speech API billed per character or per credit, glued together by your own orchestration layer and your own turn-taking logic, which is the part that is genuinely hard. Voxtral 3 does that gluing itself, on LiveKit, and sells the result at a single per-minute rate. That is a procurement argument before it is a technical one, and procurement arguments are what move enterprise deals.

The three-stage pipeline, and why it is not just a bundle

The SDK runs listening, reasoning and speaking as one streaming pipeline. Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking — the end-of-turn detection that decides whether your agent waits politely or talks over somebody. A general model, Mistral Small 4 by default, reasons and calls tools. Voxtral Speech 3 replies in the tone of the conversation so far, which is a subtler claim than it sounds: the model is given the conversational context and adjusts emotion, pacing and emphasis from it rather than from a per-call style parameter.

Mistral's claim is that no stage waits for the one before it, so each turn completes in under one second, and that the agent stops talking when the caller cuts in. Both properties are the ones that separate a demo from a phone line. A pipeline where transcription finishes before reasoning starts, and reasoning before speech begins, adds three latencies and cannot feel conversational however good each part is; barge-in is the difference between a robot reading a script and something a caller will tolerate for four minutes. Transport is SIP for phone calls and WebRTC for browser and app sessions, which is the correct pair and covers both the contact-centre case and the in-app assistant case.

What Mistral has not published is as notable as what it has. The models are in preview with no GA date and no SLA. There is no technical report for Voxtral 3 in the launch material, no parameter count for anything except Mini at 4B, and no weights or licence terms for the self-hosted path — which is a commercial conversation, not a download. And the launch post and the documentation do not fully agree with each other: the post describes TTS as covering 35+ languages while the model documentation has been read elsewhere as listing a smaller set, and the realtime model's minimum latency appears as 240 ms on the chart and 80 ms in the documentation. Neither discrepancy is fatal; both are the kind of thing that gets resolved after preview, and both are a reason to test before you standardise.

Key Features

Transcription — three tiers, one price ladder

• Voxtral Mini Transcribe 3 — $0.0035/min. A 4B-parameter batch model built for cheap, fast jobs at volume. This is the cheapest per-minute transcription in the family and the one to use for archives, call recordings and subtitle passes where a file already exists.
• Voxtral Small Transcribe 3 — $0.0040/min. The flagship transcriber, and the one Mistral says is strongest on multilingual transcription and diarization — the "who said what" problem. At 32K of context it also does translation, speech question answering, summarization and function calling, which makes it an audio-language model that transcribes rather than a pure transcriber.
• Voxtral Mini Transcribe Realtime 3 — $0.008/min. Streaming over WebSocket with end-of-turn detection, 35+ languages and latency under 500 ms. This is the tier the voice agent SDK listens with, and the tier where the accuracy trade in the next section lives.
• 50+ languages across the batch tiers, including all 24 official EU languages and a set of low-resource Asian and Indic languages. That EU list is not decoration: it is the reason a European public-sector buyer will read this release at all, because "all 24 official languages" is a procurement checkbox that "50+ languages" is not.
• Diarization on both batch models, with Mistral reporting a concatenated min-permutation WER of 18.4% for Small against 38.4% for Gemini 3.5 Transcribe on its own harness. That is the largest gap in the launch material and the one most likely to matter for meeting notes and call-centre QA — and the one Google has published no counter-number for.
• Context biasing, no fine-tuning. You hand the model key terms or free-form text and it biases recognition toward them, which lifts Mistral-reported keyphrase recall from 81.4% to 94.6%. For anyone transcribing product names, drug names, part numbers or surnames, this replaces the glossary-and-post-processing hack that most pipelines carry.
• End-of-turn detection. Listed as a capability of the realtime model and load-bearing for the SDK, because an agent that guesses wrong either interrupts or stalls.

Speech synthesis — priced by the character, shaped by the context

• Voxtral Speech 3 at $13 per million characters, in batch and realtime, on the same API. A single price for both modes is genuinely unusual — most vendors charge a premium for streaming TTS — and it means you can prototype in batch and ship realtime without a cost review.
• 35+ languages and 50+ dialects, with explicit controls for emotion, pacing and emphasis, plus contextual TTS that infers tone from the conversation, and a vocabulary library for pronunciation of names and jargon.
• 34 emotion tags and roughly 90 ms time to first audio, per the model page rather than the launch post. The latency figure is the one that makes TTS usable inside a sub-second turn; the tag count is the one that makes it expressive.
• Voice cloning from 30+ minutes of audio, and Pro Voice — also 30 minutes — which reproduces a speaker's distinctive character, emotional range and style across languages, so a profile cloned in English can speak German or Japanese in the same voice.
• Both cloning modes raise consent questions. They are available in Mistral's Studio playgrounds, which is where a great many people will first try them on the first voice they have to hand. See the privacy note below before you ship one.

The voice agent SDK — one vendor, one meter

• All four stages stream as one pipeline: listen, detect end of turn, reason, call tools, speak. Mistral claims under one second per turn and barge-in handling.
• $0.04/min with Mistral Small 4 as the reasoning model; $0.08/min if you swap in a different model. The surcharge is the honest part — it tells you the bundle is priced around Mistral's own inference margin, and it means bringing your own LLM doubles the rate.
• SIP for telephony, WebRTC for browsers and apps. Both transports, no third-party gateway required for the browser case.
• Built on LiveKit, which is open source. Convenient now, and a rewrite if you ever move off it — the orchestration, turn detection and media handling live in that layer.
• Testable in Mistral's Studio playgrounds for voice agents, transcription, generation and cloning, without writing code.

Deployment — the sovereignty angle

• EU, US and global endpoints on the Mistral API, so you can pin audio processing to a jurisdiction rather than hope.
• Self-hosting on your own infrastructure, positioned as sovereign deployment with customer control over audio retention. For a European hospital, bank or ministry, audio is the most sensitive data in the building and "it never leaves our network" is the requirement, not a nice-to-have.
• No third-party media vendor in the path under the self-hosted option — which is the specific thing that disqualifies several hosted voice-agent platforms in regulated procurement.

The Accuracy Question: Mistral's Own Chart Is the Story

Most launches give you one flattering number. This one gives you a table in which Mistral loses a row, and that row is the most useful thing in the release for anyone building a live agent.

All rows vendor-reported by Mistral as of October 20, 2026, on FLEURS across 35 languages, macro-averaged. Competitor rows carry no stated delay, so they are not a controlled latency-matched comparison. No independent reproduction of the realtime chart existed at the time of writing.

Three readings, in ascending order of how much they should change your plan.

• The latency tax is real and it is priced, not hidden. Same model, same audio, same 35 languages: 4.0% as a file, 7.4% at a quarter second. That is roughly 85% more word errors for immediacy. Every streaming ASR vendor faces this curve; publishing it with five points on it is more honesty than the category usually gets. If your use case can tolerate 960 ms, you buy back 1.4 points of accuracy for free.
• Voxtral is mid-pack in realtime, not last. It beats Scribe v2 Realtime at 8.7% and GPT Live Transcribe at 8.4%, and loses to Gemini 3.5 Realtime at 4.2% and Microsoft AI Transcribe 2 Streaming at 5.8%. Anyone reporting this as "Mistral's realtime model is bad" is overreading one row; anyone reporting it as "state of the art" is ignoring the same row. The accurate statement is that Mistral's realtime tier is roughly a second-tier accuracy option with a first-tier price, and the price is what compensates.
• The comparison is not latency-matched, and that cuts both ways. Mistral's chart puts a 240 ms Voxtral row next to competitor rows with no stated delay. If Gemini's 4.2% is measured at a more generous delay, the gap is partly an artefact of the chart; if it is measured at a comparable one, Gemini is simply better at streaming. Mistral chose not to say, and the absence of the number is the tell. Test it on your own audio before you believe either side.

Pricing: Four Meters, One Invoice

Every figure below is in US dollars, checked on October 22, 2026, against Mistral's launch pricing table and model documentation. There is no peak/off-peak split and no long-context surcharge.

Worked example 1: a support line doing 1,000 minutes a day

This is the headline case, and it is the one worth internalising because it is where a per-minute rate stops feeling small. A support line handling 1,000 minutes of conversation per day — about 17 hours of talk time, so a handful of concurrent agents — on the voice agent API at $0.04 per minute:

• $0.04 × 1,000 minutes = $40 per day.
• $40 × 30 = about $1,200 per month.
• $1,200 × 12 = about $14,400 per year, on the voice pipeline alone.
• Swap in your own reasoning model at $0.08 per minute and the same traffic costs $80/day, about $2,400/month — the single most expensive decision in the architecture, and it is a config flag.

Now add the part the rate card does not include. The $0.04 covers listening, reasoning and speaking. It does not cover the phone line. A production support line also pays for SIP numbers, typically a few dollars per number per month, and inbound PSTN minutes from a carrier. At a common US inbound rate of roughly $0.0085 per minute, 30,000 minutes a month adds about $255; your carrier's rate will differ and you should check it rather than budget on ours. Add recording storage, the CRM writes, and the monitoring you will want on a line that customers call, and the honest all-in figure for this workload lands meaningfully above $1,200 and comfortably below $2,000. Budget the $1,200 and the $255 separately, because they are different invoices with different owners.

Worked example 2: transcription only, same 1,000 minutes a day

If you are transcribing recorded calls rather than talking to anyone live, the arithmetic is an order of magnitude smaller, and this is the comparison that makes the family look cheap:

• Mini at $0.0035/min: $3.50/day → about $105/month.
• Small at $0.0040/min: $4.00/day → about $120/month.
• Realtime at $0.008/min: $8.00/day → about $240/month — and note that streaming costs 2.3× the batch Mini rate for a model that is, on Mistral's own numbers, less accurate at low latency.
• TTS at $13/1M characters: at 900 characters per minute of speech, 1,000 minutes costs $11.70/day, about $351/month; at 1,000 characters per minute it is $13.00/day, about $390/month.

That last line is the conversion that matters most, and almost nobody publishes it. Mistral quotes speech in characters; its competitors quote in credits, characters or minutes depending on the vendor. To compare at all you need a character rate for speech, and it is not a constant — it depends on language, speaking speed and punctuation density. Our working assumption for English is 900 to 1,000 characters per minute, which puts Voxtral Speech 3 at about $0.0117 to $0.013 per minute of audio. That is our arithmetic, not Mistral's published figure, and you should redo it against your own scripts.

The bundle is not the sum of the parts — and that is informative

If you bought the components separately for a pipeline you orchestrate yourself, a minute of speech-to-speech would cost $0.008 for realtime transcription plus about $0.012 for the speech it generates — roughly $0.020 per minute before any language model. The agent API is $0.04. So the bundle carries a premium of about 2× the component cost at the same volume, about $690 a month on our 30,000-minute example, and what that buys is the LiveKit orchestration, turn detection, barge-in and media handling you would otherwise build and operate. Whether that is good value depends entirely on whether you have engineers who want to own a real-time media pipeline. If you do, building is cheaper. If you do not, $0.04 is a fair price for not finding out how hard barge-in is.

Is It Free?

No, and it is worth being blunt about it because the family's low rates invite the assumption.

• There is no permanent free tier that we could find. All four models are in public preview, which governs stability and price risk rather than cost — Mistral publishes a rate card, not an allowance. Every call is billed at the rates above.
• The playgrounds are the free door. Mistral's Studio has playgrounds for voice agents, transcription, speech generation and voice cloning, and they are the right place to form an opinion about output quality before you pay for one. Read the retention terms before you upload a customer recording or clone a real person's voice into a free surface.
• Trial credits may exist; confirm in your console. Mistral's console has historically issued trial credits to new accounts, but the amount and conditions are not stated in the Voxtral 3 launch material. Do not budget on a number from a review — log in and look.
• Self-hosting removes a licence fee, not a bill. A 4B transcriber plus a TTS model plus a reasoning model is a serving fleet. The GPU hours, the autoscaling, the on-call rota and the monitoring all survive the move to your own hardware, and the licence terms for the self-hosted path are a negotiation rather than a download because no weights were published with this release.
• Preview pricing is not a commitment. Public preview means Mistral has not promised that $0.0035 or $0.04 survives GA. Budget at today's rate and treat any increase as a known risk rather than a surprise.

Practical read: prototype in the playground, run your own accuracy test on $5–10 of API spend — that is roughly 1,400 minutes of batch Mini or 125 minutes of full voice agent — and then build the business case on the published rates plus telephony.

Pros & Cons

5 Best Mistral Voxtral 3 Alternatives

Comparison Table

Prices approximate, in US dollars, checked October 22, 2026. Where a vendor does not publish a directly comparable figure we say "check official pricing" rather than guess. Latency and WER in the Gemini and Mistral rows come from Mistral's own chart, not from Google's. Whisper's $0 is a licence cost, not a total cost: GPU hours and operations are yours.

Which One Should You Choose?

• Call centre or support line — the strongest reason to buy this release. $0.04/min all-in, under a second per turn, barge-in, SIP for telephony, end-of-turn detection, and one vendor to escalate to. Budget $40/day per 1,000 minutes plus telephony and recording storage, and test the realtime WER on your own call audio at both 240 ms and 960 ms before you commit — the second setting may be accurate enough and is free to choose. If your calls are accented, noisy or code-switched, run 40 minutes of the worst audio you own through it before you sign anything.
• Multilingual subtitles and archive transcription — buy the batch tiers and stop reading. $0.0035 and $0.0040 per minute, 50+ languages including all 24 EU official languages, diarization at 18.4% cpWER, context biasing that lifts keyphrase recall to 94.6%, and independent corroboration at 2.1% WER against Gemini's 2.6%. This is the best-value part of the release by a wide margin, and 1,000 minutes a day costs about $105 a month. Use Small when speakers overlap, Mini when they do not.
• Cloned narration and localisation — Pro Voice is the feature to pilot. 30 minutes of reference audio, character and emotional range carried across 35+ languages, 34 emotion tags and roughly 90 ms to first audio. Compare against ElevenLabs on the same script before deciding, because naturalness on long-form narration is where ElevenLabs has historically been strongest and where a per-character price difference of 13 percent is not the deciding factor. If your script is a 40-minute audiobook, pay for the voice; if it is 400,000 short agent utterances, pay for the rate.
• Data sovereignty or regulated audio — this is the shortlist entry. EU, US and global endpoints, a self-hosted path with no third-party media vendor in the audio path, and a European vendor. Very few competitors in this table can say all three. Price the GPU bill before you believe self-hosting is cheap, and get the licence terms in writing, because none were published with the release.
• Pure budget — the batch tiers win, the realtime tier is a judgement call. At $0.0035 and $0.0040 per minute for transcription, Voxtral is at the bottom of the market. At $0.008 for realtime, you are paying 2.3× the batch rate for the least accurate model in the family, and Gemini's live SKU is roughly $0.009 with better realtime accuracy — so the budget argument for realtime is weak and the argument for batch is overwhelming. Our LLM API cost guide covers the levers that move a bill more than any headline rate.
• You already have a pipeline — buy components, not the bundle. Realtime STT at $0.008 plus TTS at about $0.012 is roughly $0.020 per minute before the language model, against $0.04 for the all-in agent. If you have engineers who want to own a real-time media pipeline, that 2× is real money. If you do not, it is the cheapest way to avoid learning how hard barge-in is.

Why Not Just Use ElevenLabs?

It is the right question, because ElevenLabs is the name most buyers reach for and, for a lot of jobs, it is still the correct answer. The two products are not really the same shape, though, and the difference is best stated plainly.

ElevenLabs is a voice product. Voxtral 3 is a voice stack. ElevenLabs sells you the best-sounding synthetic voice in the market plus the tooling around it — a library, dubbing, a studio editor, per-language voices. If your deliverable is a narrated asset, an audiobook, a localised video or a branded agent voice, and the constraint is "does it sound like a person," that is the product you want, and the subscription from about $6 to about $990 a month is a rounding error against the cost of a re-record. Mistral is selling infrastructure: transcription at $0.004 a minute, synthesis at $13 per million characters, and a $0.04-a-minute pipeline that also listens, reasons and calls tools. It is cheaper per unit and it is not trying to be a studio.

Three places Voxtral genuinely wins. First, volume economics: at 30,000 minutes a month, per-minute pricing beats credit pricing, and the gap widens with scale because credits do not get cheaper as fast as minutes do. Second, one contract instead of two: with ElevenLabs you still need a transcriber, which means a second vendor, a second bill and a second place for the transcript to go wrong — and Mistral's own batch quality is now independently corroborated. Third, sovereignty: EU endpoints and a self-hosted path with no third-party media vendor in the audio path, which ElevenLabs does not offer and which decides the deal outright for a European regulated buyer.

Three places ElevenLabs wins and you should not pretend otherwise. Voice quality on long-form expressive speech is still the category benchmark, and 34 emotion tags is not the same thing as a decade of voice data and a dubbing workflow built on top of it. The product is mature — an ElevenLabs integration does not carry public-preview risk on the endpoint your customers hear. And the voice library is a genuine moat: if your requirement is "a warm Scottish female voice for this brand," you can audition it in an afternoon rather than cloning one and hoping.

The honest recommendation. If you are building an agent that talks to customers on a phone line and you care about cost per minute and where the audio is processed, Voxtral 3 is the better fit and the realtime accuracy trade is the price of admission — test it on your own audio and consider 960 ms. If you are producing audio for humans to listen to and enjoy, ElevenLabs is worth the premium and the per-character comparison is the wrong frame entirely. And if you need both, which is common, buy them from both vendors and stop pretending one product should be excellent at both jobs. The full product comparison is in our ElevenLabs review.

Privacy & Consent Note

Voice is the most sensitive data most companies handle without a policy for it, and this release makes two things easier that were previously awkward: processing audio inside a jurisdiction you choose, and generating a synthetic version of a specific person's voice. The second is where the risk sits.

• Voice cloning needs documented consent, not verbal assent. Voxtral Speech 3 clones from 30+ minutes of audio and Pro Voice reproduces character, emotional range and style across languages from the same 30 minutes. Thirty minutes is easy to obtain and hard to justify afterwards if you did not ask. Get written consent that names cloning specifically, from the speaker, for the specific use — and keep it, because the burden of proving consent will be yours.
• A voice recording is personal data, and a voiceprint can be special-category data. Under GDPR, audio of an identifiable person is personal data; where a voice model is used to identify someone, it can be treated as biometric data, which carries stricter conditions and generally requires explicit consent or another Article 9 basis. This is not legal advice and the analysis varies by jurisdiction and by what you do with the model — but "we recorded it, so we can clone it" is not a defensible position anywhere in the EU.
• Disclose synthetic speech where the law or decency requires it. The EU AI Act's transparency provisions for deepfakes, plus publicity-rights and consumer-protection law in the United States, mean an agent that sounds like a real, named person needs a disclosure — and an agent that imitates a celebrity needs a lawyer. Default to telling the caller they are speaking to a synthetic voice; it costs nothing and it is increasingly required.
• Retention is the question to ask sales. Mistral positions the self-hosted path as customer control over audio retention, which is a real advantage — but retention on the hosted API is governed by Mistral's terms and data-processing agreement, and the launch post does not summarise them. Ask for the DPA, ask what happens to audio after inference, and ask whether recordings are used for training. Do not infer an answer from a sovereignty headline.
• The playgrounds are a live surface. Voice agents, transcription, generation and cloning are all testable in Mistral's Studio without writing code. That is excellent for evaluation and a bad place to upload a customer call recording or a colleague's voice sample. Treat the playground as public infrastructure until a contract says otherwise.
• Third-party audio is the sharp edge. Most cloning incidents are not malicious; they are a team testing a feature on a podcast clip, a sales call or a voice actor's demo reel. Write the rule before the temptation arrives: consenting voices only, named owner, deletion date.

Two mitigations that cost nothing: pass references rather than raw recordings where your design allows, and keep the evaluation environment and the production environment under separate keys so a playground experiment cannot become a production data flow by accident.

Frequently Asked Questions