On 20 October 2026, Mistral AI unveiled Voxtral 3, its third generation of voice models. The announcement boils down to one promise: a voice agent that listens, thinks, acts and replies in under a second, in dozens of languages, for 4 US cents a minute. Enough to run a switchboard or an after-sales service day and night, with an argument that American competitors cannot make in quite the same way: hosting in Europe.
The price is real and low. So is the technical progress, and an independent measurement confirms it. But three things go unnoticed in most write-ups of the announcement. None of the four new models can be downloaded, a first for the range. Several “first places” do not hold up when Mistral’s charts are read to the end. And European hosting says nothing, on its own, about how long your conversations are kept.
Voxtral 3: what Mistral launched on 20 October
Voxtral 3 is not one model but a set of building blocks, sold separately or assembled into an agent.
| Building block | Role | Announced price |
|---|---|---|
| Voxtral Small Transcribe 3 | Most accurate transcription in the range, speaker identification | $0.004 a minute |
| Voxtral Mini Transcribe 3 | Batch transcription, 4 billion parameters | $0.0035 a minute |
| Voxtral Mini Transcribe Realtime 3 | Live transcription, the agent's "ear" | $0.008 a minute |
| Voxtral Mini TTS 3 | Speech synthesis, the agent's "voice" | $13 per million characters |
| Voice agent API with Mistral Small 4 | The complete agent, reasoning included | $0.04 a minute |
| Voice agent API with "a larger model" | The same agent, with a more powerful reasoning model | $0.08 a minute |
In Mistral’s documentation, all four models are marked “Public Preview”: they are available, but as a preview. They are offered through Mistral’s API, on European, US and global endpoints, or for deployment on the customer’s own infrastructure. To run the whole stack on your own servers or in a private cloud, the announcement points to a conversation with the sales team.
Mistral is also launching “Pro Voice”, a professional voice adaptation that reproduces a person’s timbre, emotional range and way of speaking, across the supported languages, from 30 minutes of recording.
AI voice agent: how Mistral’s works
An AI voice agent is a program that can hold a spoken conversation: it understands what it is told, decides what to do, acts in other software if needed, then replies out loud. At Mistral, these functions are split across three separate models, chained together continuously.
- The ear. Voxtral Mini Transcribe Realtime 3 transcribes speech as it comes and detects when the other person has finished speaking, combining acoustic cues with the meaning of the sentence.
- The brain. A general-purpose language model, Mistral Small 4 in the base offer, understands the request and calls tools: verifying a caller’s identity, checking a bill, booking an appointment, searching the web or the company’s own databases.
- The voice. Voxtral TTS replies, in a tone matched to the conversation.
According to Mistral, no stage waits for the previous one to finish, so each turn completes “in under one second”, and the agent stops talking when it is interrupted. The whole thing runs over the phone (the SIP protocol) or in a browser or app (WebRTC), and connects to existing switchboards and contact-centre software.
Two details matter. First, the development kit is built on LiveKit, a third-party real-time communication technology, rather than on an in-house component. Second, because each stage is a separate model, you can change the brain, the voice or the language without rewriting the agent. That explains the two price points: $0.04 a minute with Mistral Small 4, $0.08 with “a larger model”. The announcement does not say which one, and nothing in it indicates that it is Mistral Large 4.
How much an AI voice agent costs at Mistral
At $0.04 a minute, a 10-minute call costs $0.40, and an hour of conversation $2.40. This rate covers the agent, reasoning included in the Small 4 version. The price table, however, mentions neither a phone number nor call charges: according to Mistral, the agent connects “to the telephony” the company “already” uses. The phone subscription and the integration with in-house software therefore have to be budgeted separately.
Two benchmarks help put this price in context. The ear alone, realtime transcription, costs $0.008 a minute, one fifth of the price of the complete agent. At ElevenLabs, one of the main players in synthetic voice, the pricing page shows $0.08 a minute for its voice agent pipeline, and $0.08 plus language model usage for its models that turn speech directly into speech. Mistral’s base offer is therefore listed at half the price. These are list prices, and quality is not the same everywhere, as we will see below.
Another point rarely mentioned: the price of a minute of transcription is going up from one generation to the next.
| Criterion | Voxtral Transcribe 2 (February 2026) | Voxtral 3 (October 2026) | Change |
|---|---|---|---|
| Batch transcription (Mini model) | $0.003 a minute | $0.0035 a minute | +17% |
| Realtime transcription | $0.006 a minute | $0.008 a minute | +33% |
| Languages transcribed | 13 | 50 | Almost four times as many |
| Licence of the realtime model | Apache 2.0, weights published | "PREMIER", weights not published | Closed |
The increase comes with a real jump in coverage, from 13 to 50 languages. But it runs against the idea of voice AI that costs next to nothing: at equal volume, the agent’s ear costs a third more than in February.
Is Voxtral 3 open source? No, and that is a first for the range
Mistral’s official model catalogue gives each model a licence label. The label on all four Voxtral 3 models is “PREMIER”, the label for models without an open licence. As of 28 October 2026, none of them appears on Mistral’s Hugging Face page, where the company publishes its downloadable models.
| Generation | Model | Licence (Mistral catalogue) | Free commercial use |
|---|---|---|---|
| Voxtral (July 2025) | Voxtral Small, Voxtral Mini | Apache 2.0 | Yes |
| Voxtral Transcribe 2 (February 2026) | Voxtral Mini Transcribe Realtime | Apache 2.0 | Yes |
| Voxtral Transcribe 2 (February 2026) | Voxtral Mini Transcribe 2 | PREMIER | No, API access |
| Voxtral TTS (March 2026) | Voxtral TTS | CC BY-NC 4.0 | No, non-commercial use only |
| Voxtral 3 (October 2026) | All four models | PREMIER | No, API access or by contract |
Voxtral 3 is therefore the first Voxtral generation in which no model can be downloaded. It should be said fairly: the announcement promised nothing else, it mentions neither weights nor licence. But the practical consequence is clear. “Self-deployment on your own infrastructure”, presented as a guarantee of sovereignty, goes through a commercial agreement with Mistral: you cannot download Voxtral 3 and run it without Mistral.
The agent’s brain, on the other hand, is open: Mistral Small 4 is under the Apache 2.0 licence. As for Mistral Large 4, presented on 6 October as an open-weight model, its model card still shows “Coming soon” on 28 October for both the weights and the licence, as we noted in our article on Mistral Large 4. The underlying trend is not specific to Mistral: it ties in with the question of who, today, actually releases their best models openly.
For anyone looking for an open Voxtral, the best option therefore remains a 2025 model: Voxtral Small, still under Apache 2.0, measured at 2.8% errors by Artificial Analysis, against 2.1% for Voxtral Small Transcribe 3.
What Mistral’s charts say when you read them in full
The Voxtral 3 announcement contains about ten charts. Read in full, they are more nuanced than the text around them.
| What Mistral claims | What its own chart shows |
|---|---|
| Small Transcribe 3 outperforms Microsoft AI Transcribe 2 (model card) | FLEURS test, 45 languages: a tie to the tenth of a point, 4.0% errors with the language specified and 4.1% with the language detected, for both. On the Artificial Analysis index reproduced in the announcement: Microsoft 2.04%, Voxtral 2.10%. |
| "State-of-the-art" transcription | Third on the Artificial Analysis index reproduced by Mistral, behind Alibaba's Fun-Realtime-ASR-preview (1.73%) and Microsoft (2.04%). |
| "High-quality" live transcription under 500 milliseconds | FLEURS test, 35 languages: 6.4% errors at the 480 ms setting, against 4.2% for Google's Gemini 3.5 Realtime Transcribe. |
| Best speaker identification | Correct: 18.4%, first ahead of Pyannote (19.0%) and Microsoft (20.3%). But about 18 words in 100 are still mistranscribed or attributed to the wrong person. |
| Speech synthesis cheaper than OpenAI, Deepgram, ElevenLabs Flash and Gemini Flash | On numbers and proper names in customer service: 79.7% faithful clips, against 98.3% for Eleven v4, six times more expensive. |
Transcription: a tie presented as a win. The announcement text states that Voxtral Small Transcribe 3 “outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe”. Microsoft is missing from that sentence, even though it appears on the chart at the same level as Voxtral. The model card goes further and lists Microsoft AI Transcribe 2 among the models outperformed. On transcription accuracy, that is not true on either of the two Mistral charts that compare them: a tie on one, a slight lag on the other. Voxtral only clearly beats Microsoft on speaker identification.
Realtime: the agent’s ear behind Google on Mistral’s own test. This is the model that matters most for a voice agent, since it is the one listening during the call. The chart shows it at four delay settings: 5.7% errors when waiting 2.4 seconds, 6.0% at 960 ms, 6.4% at 480 ms and 7.4% at 240 ms. The faster you want to reply, the more mistakes it makes. At the setting that matches the “under 500 ms” promise, it makes about 50% more errors than Google’s realtime model (4.2%). The chart does not say at what delay Gemini was measured, which limits the comparison. At the same setting, Voxtral beats OpenAI’s GPT Live Transcribe (8.4%) and ElevenLabs’ Scribe v2 Realtime (8.7%), but stays behind Microsoft’s realtime version (5.8%).
Speaker identification: first, but far from flawless. This is the function that tells you, in a meeting, who said what. The measure used, cpWER, counts both mistranscribed words and words attributed to the wrong person, on recordings of meetings and phone calls. Voxtral Small Transcribe 3 gets the best score, 18.4%, against 30.9% for the previous generation (Voxtral Mini Transcribe 2) and 38.4% for Gemini 3.5 Transcribe. The progress is spectacular. But an automatic meeting summary would still contain, on these tests, about one word in five or six to correct.
Speech synthesis: cheaper, not more faithful on what matters in customer service. The Voice Code Bench test, designed by the company BeSimple AI, targets precisely the typical failures of customer service: reading out a case number, an email address, a proper name. On 300 English sentences, Voxtral TTS produces 79.7% faithful clips. ElevenLabs’ Eleven v4 reaches 98.3%, and its Turbo version 97.5%. Mistral presents this result as “quality per dollar”, where its model, at $13 per million characters against $40 and $80, looks well placed. In plain terms, on these sentences, Mistral’s voice gets it wrong about once in five, ElevenLabs’ fewer than twice in a hundred. On other tests, Voxtral matches or beats it: first on controlling the pronunciation of rare words (86.8% against 68.2% for Eleven v4), level with Eleven v4 on average quality across 11 languages (3.16 against 3.17 out of 5).
One caveat applies to all these speech synthesis tests: they are scored by an AI model, Google’s gemini-3.8-flash, and three of the four are flagged by Mistral as internal evaluations.
Voxtral vs Whisper: the gap is real
Whisper large-v3, OpenAI’s open transcription model, remains the free reference that many people use. On the FLEURS test in the announcement (45 languages), it makes 14.9% errors when the language is specified and 16.0% when it has to guess, against 4.0% and 4.1% for Voxtral Small Transcribe 3: almost four times as many. On the Artificial Analysis index reproduced in the announcement, the gap narrows but remains twofold, 4.07% against 2.10%.
The comparison does have a limit, though: Whisper can be downloaded for free, Voxtral 3 cannot. For local transcription, on your own computer, the open alternative to Whisper remains the 2025 Voxtral Small.
Independent measurements: real progress, not dominance
Artificial Analysis, which evaluates AI models independently of their publishers, has added Voxtral 3 to its leaderboards. Its transcription index combines three datasets, one of which, made up of conversations with agents, counts for half.
In batch transcription, as of 28 October 2026, Voxtral Small Transcribe 3 is third with 2.1% errors, behind Alibaba’s Fun-Realtime-ASR-preview (1.7%) and Microsoft’s MAI-Transcribe-2 (2.0%). The latter is also cheaper in this leaderboard: $1.67 per 1,000 minutes, against $4 for Voxtral Small Transcribe 3. Further down are ElevenLabs’ Scribe v2 and Voxtral Mini Transcribe 3 (2.2%), Gemini 3.5 Transcribe (2.6%) and OpenAI’s GPT Transcribe (3.3%). Compared with Voxtral Mini Transcribe 2 (3.6%), the progress is clear.
In realtime, the measurement that directly concerns a voice agent, the leaderboard shows this:
| Model (realtime) | Error rate | Time to final transcript |
|---|---|---|
| MAI-Transcribe-2-Streaming (Microsoft) | 2.51% | 0.13 s |
| Grok Voice Transcribe 2.0 | 2.73% | 0.49 s |
| Voxtral Mini Transcribe Realtime 3, 960 ms setting | 3.34% | 1.12 s |
| Voxtral Mini Transcribe Realtime 3, 480 ms setting | 3.57% | 0.75 s |
| Scribe v2 Realtime (ElevenLabs) | 3.59% | 0.14 s |
| GPT Live Transcribe (OpenAI) | 3.92% | 0.81 s |
| Gemini 3.5 Transcribe Live (Google) | 4.00% | 0.40 s |
| Voxtral Mini Transcribe Realtime (February 2026, open) | 5.24% | 0.68 s |
Three lessons. First, Voxtral 3’s ear improves sharply on the open February version (3.57% errors against 5.24%) and sits at ElevenLabs’ level, ahead of OpenAI and Google. Second, the order depends on the test: on Mistral’s multilingual chart, Google came out ahead; on the Artificial Analysis index, it is the other way round. For a French-speaking contact centre, neither replaces a trial on its own calls.
Finally, latency. At the 480 ms setting, Artificial Analysis measures 0.75 seconds between the end of speech and the final transcript, against 0.13 to 0.14 seconds for Microsoft and ElevenLabs. Mistral’s “under 500 ms” refers to a delay setting, not to this measurement. That does not necessarily contradict the promise of a turn in under a second, since Mistral says its stages work in parallel, but the margin is thin. As of 28 October, we found no independent measurement of a complete agent, covering listening, reasoning, tools and voice.
50 languages? What the agent hears, and what it can say
According to the announcement, Voxtral 3 understands 50 languages in batch transcription, including the 24 official languages of the European Union and many Asian and Indian languages. That is a considerable leap from the 13 languages of February.
But the voice agent does not rely on that model. Its ear, the realtime model, covers “over 35 languages” according to the announcement text and “40+” according to its model card, with no single list. And its voice speaks “35+ languages and 50+ dialects”, with no published list.
Ready-to-use voices, the ones you can use without recording anyone, exist in 18 languages. Among them are seven official EU languages (Dutch, English, French, German, Greek, Italian, Spanish), plus Portuguese in its Brazilian variant only. Polish, Swedish, Czech and Romanian have no ready-made voice, even though Mistral’s sample library lets you hear some of them. The French voice on offer is from France: no Belgian, Swiss or Québécois variant ready to use. The “24 official EU languages” therefore apply to what Voxtral 3 hears, not to what it says.
Sovereign: what European hosting guarantees, and what it does not
European hosting is real: Mistral offers an endpoint in Europe, alongside US and global endpoints. You still have to pick the right one. But where data is processed is only part of the question. The other part is how long it is kept.
According to Mistral’s privacy policy in force since 3 September 2026, inputs and outputs of its API are kept for “thirty (30) rolling days” to monitor abuse, except for specific APIs, and unless zero data retention is activated. For the Agents API, they are kept until the account is terminated.
Zero data retention exists, but with conditions. According to Mistral’s documentation, it is limited to paid plans and to “stateless” calls, those that keep no memory from one exchange to the next. It has to be requested with a legitimate reason, and Mistral grants or refuses it at its discretion. Transcription and speech generation are eligible; agents and conversations are excluded. The Voxtral 3 voice agent API appears on neither list.
For a company that would route customer calls through a voice agent, sometimes with banking or health data, the question to ask Mistral before any deployment is therefore simple: which retention regime applies to this API? It is the same reflex we recommended for finding out who processes your data when a bank or a health service uses AI. Sovereignty has other dimensions, which show up in Mistral hosting a Chinese model: a European infrastructure does not tell you everything about what runs on it.
As for deployment on your own servers, it is possible, but by contract with Mistral, since the models are not published. It is a real option for controlling data. It is not independence from the publisher.
Voice cloning: the obligation to say “I am an AI”
With Pro Voice, a company can make its agent speak with the voice of a real person, from 30 minutes of recording. The announcement presents it as the way to reply in a brand’s own voice. The more human the voice, the more one rule matters.
Article 50 of the EU Artificial Intelligence Act requires that people who interact directly with an AI system be informed that they are talking to an AI, unless this is obvious to a reasonably observant person. This information must be given clearly, at the latest at the first interaction. Under Article 113 of the regulation, this obligation has applied since 2 August 2026. The obligation falls on the provider of the system, which may be the company itself if it builds its agent and puts it into service under its own name. A voice agent with a human voice must therefore say so at the start of the call. The text’s other deadlines are detailed in our article on what changes with the AI Act.
What will need checking
- An independent test of a complete agent. The real duration of a turn, errors on case numbers and addresses, how interruptions are handled: none of this has yet been measured outside Mistral.
- The model behind the $0.08 rate. Mistral refers to “a larger model” without naming it.
- The end of the preview. The models are in “Public Preview”: prices and performance may still change.
- Any published weights. None for Voxtral 3, and still none for Mistral Large 4, three weeks after its announcement.
- The retention regime of the voice agent API, absent from the zero data retention documentation.
- The full list of speech synthesis languages, announced as “35+”.
Voxtral 3 marks clear progress on the previous generation, and its entry price is low for a complete voice agent. Our ranking of audio and voice AI already tracks it. But a voice agent is not judged by its best chart. It is judged by the case number it reads correctly to a customer in a hurry, by the second it takes to reply, and by what happens to the conversations once the call is over. On all three points, the figures published to date argue for testing before signing.