Models
What Is Voxtral 3? Mistral’s Voice Agent Models and SDK, Explained
Voxtral 3 is Mistral’s new set of speech models plus a voice agent SDK, launched October 20, 2026. The four models, prices, languages and how to try it.
Voxtral 3 is Mistral’s answer to a simple question: can an AI talk on the phone like a person, quickly and in many languages? On October 20, 2026, the French AI company released a new round of speech models and a kit for building voice agents, all under the Voxtral 3 name.
If you saw "Mistral Voxtral 3" in the news today and want to know what it is, what it costs and whether you can use it, this page covers it in plain English.
What is Voxtral 3?
Voxtral is Mistral’s family of audio models. The first Voxtral models came out in July 2025 and could understand speech. Voxtral Transcribe 2 followed in February 2026, and a text-to-speech model, Voxtral TTS, arrived in March 2026.
Voxtral 3 (also typed "Voxtral3" or "Mistral Voxtral 3") is the third generation, and the big change is the goal. Mistral’s announcement says the release is about building "agents that can listen, reason, speak, and act." In other words, it is less about turning audio into text and more about running a full spoken conversation where the AI can also do things, like look up an order or book an appointment.
Voxtral 3 models: what is in the release
Mistral released four models and one developer kit. Here they are side by side, with prices from Mistral’s own announcement.
| Part | What it does | Languages | Price |
|---|---|---|---|
| Voxtral Small Transcribe 3 | Most accurate speech-to-text, labels who said what | 50+ | $0.0040 per minute |
| Voxtral Mini Transcribe 3 | Cheaper 4B-parameter speech-to-text for batch jobs | 50+ | $0.0035 per minute |
| Voxtral Mini Transcribe Realtime 3 | Live speech-to-text in under 500 ms, knows when you stop talking | 35+ to 40+ (Mistral lists both) | $0.008 per minute |
| Voxtral Speech 3 | Text-to-speech with control over emotion, pace and emphasis | 35+ languages, 50+ dialects | $13 per 1M characters |
| Voice agent SDK | Joins listening, thinking and speaking into one live pipeline | n/a | Pay for the models it uses |
To put the prices in normal terms: one hour of transcription with Voxtral Mini Transcribe 3 works out to about 21 cents, and one hour with the realtime model to about 48 cents.
How does the Voxtral 3 voice agent work?
Think of a phone call with three people passing notes. One listens, one thinks, one speaks. The Voxtral 3 SDK does exactly that, but in a single stream:
- Listen: Voxtral Mini Transcribe Realtime 3 turns your voice into text as you talk and guesses when you have finished, using both sound and meaning.
- Think and act: Mistral Small reads the text, decides what to say and can call tools, such as a web search or a company’s own systems.
- Speak: Voxtral Speech 3 reads the answer out loud.
Mistral says each turn finishes in under a second, and the agent stops talking if the caller interrupts, the way a polite person would. The SDK is built on LiveKit, an open-source tool for live audio and video. It connects to phone lines over SIP and to browsers and apps over WebRTC. Because each step is a separate model, a developer can swap in a different brain or voice without rebuilding the agent.
Voice cloning and Pro Voice
Voxtral Speech 3 comes with ready-made voices in 18 languages. Businesses can also clone a voice, and Mistral’s Pro Voice feature adapts a speaker’s character, emotional range and style across languages from about 30 minutes of recorded audio. Developers can tell the model how to say a line with explicit settings for emotion, pacing and emphasis, or let it guess the mood from the text.
Voice cloning raises fair questions about consent and misuse. Crypto Briefing’s launch story flagged this as one of the open issues, and Mistral’s announcement does not describe safeguards in detail.
Voxtral 3 benchmarks
All numbers below are Mistral’s own, published on launch day:
- Transcription accuracy: on the FLEURS test across 45 languages, Mistral’s chart puts Voxtral Small Transcribe 3 at a word error rate of about 4%, compared with about 16% for OpenAI’s older Whisper large-v3. Lower is better.
- Who said what: on speaker labeling across five public tests, Mistral reports an 18.4% error rate for Voxtral Small Transcribe 3 versus 38.4% for Gemini 3.5 Transcribe.
- Cost: Mistral says Voxtral Speech 3 is cheaper to run than comparable voices from OpenAI, Deepgram, ElevenLabs Flash and Gemini Flash.
Independent tests will tell whether those gaps hold up. Speech models also vary a lot by accent and background noise, so test with your own recordings.
Is Voxtral 3 free or open source?
Not in the way earlier Voxtral models were. The July 2025 Voxtral models were open-weight under Apache 2.0. For Voxtral 3, Mistral’s announcement lists access through its API (with European, US and global endpoints), testing in its Studio playgrounds, an EU-hosted option and self-deployment on your own servers for companies that contact Mistral. It does not mention free downloadable weights. Some articles call it an "open-weight pipeline," but we could not confirm that from Mistral’s own page, so treat it as unconfirmed.
OpenRouter, a site that resells many AI models through one API, also announced Voxtral 3 support on launch day.
How to try Voxtral 3
- Sign in to Mistral’s developer platform and open the Studio playground for voice agents, transcription or speech.
- Upload a short recording or type a line of text to hear Voxtral Speech 3.
- For a real agent, follow Mistral’s voice agent SDK guide, which runs on LiveKit.
Mistral did not say Voxtral 3 is inside its Le Chat app yet, so for now it is a developer tool.
Voxtral 3 vs ElevenLabs and OpenAI voice
The voice agent race got crowded this week. ElevenLabs recently launched Eleven v4 and Eleven v4 Turbo for fast, expressive speech, and OpenAI sells its own realtime voice API. Mistral’s pitch is a full stack from one company, low prices, many languages including all 24 official EU languages, and the option to keep data in Europe. If you run a call center in Europe, that last point may matter more than any benchmark.
What this means for you
Voxtral 3 is a building block. Over the next months, many of the "press 1 for billing" phone menus you know will be swapped for AI agents that listen, look things up and answer in your language. That is agentic AI moving from chat windows to phone calls.
You do not have to wait for your bank to build one. Agentic AI already gives you a team of expert agents on iPhone and iPad for research, planning and writing, so you skip the model shopping and just ask. Mistral also shipped the Timestral forecasting models last week, another sign that AI is splitting into specialists.
Quick answers
What is Mistral Voxtral 3? A release of three speech-to-text models, one text-to-speech model and a voice agent SDK from Mistral, launched on October 20, 2026, for building AI that talks with people in real time.
Is Voxtral 3 free? No. It is a paid API. Transcription costs $0.0035 to $0.008 per minute and Voxtral Speech 3 costs $13 per million characters, per Mistral.
How many languages does Voxtral 3 support? The transcription models cover 50+ languages. Voxtral Speech 3 covers 35+ languages and 50+ dialects.
Is Voxtral 3 open source? Mistral’s announcement does not mention open weights for Voxtral 3. It offers an API, an EU-hosted option and self-deployment through Mistral sales.
Can Voxtral 3 make phone calls? Yes, developers can connect its voice agents to phone lines over SIP, as well as to apps and browsers over WebRTC.
Not sure what words like SDK or speech-to-text mean? Our agentic AI glossary explains them in plain English.
Sources: Mistral: Voxtral 3 · Mistral AI on X · Crypto Briefing · OpenRouter on X · TechCrunch: first Voxtral (2025) · ElevenLabs: Eleven v4
