TL;DR: Mistral Voxtral 3 launched on October 20, 2026 as a new family of speech-to-text and text-to-speech models plus a voice agent SDK built on LiveKit. Mistral says transcription now covers 50+ languages, speech output covers 35+ languages and 50+ dialects, and each conversational turn completes in under a second. Prices start at $0.0035 per minute for batch transcription, $13 per million characters for speech, and $0.04 per minute for a full voice agent running on Mistral Small 4. It is available through Mistral’s API or for self-hosting.
Mistral Voxtral 3 is the French AI lab’s attempt to sell the whole voice stack, not just a transcription model. The release bundles listening, reasoning, tool calling and speaking into one real-time pipeline aimed at contact centers, service desks, virtual assistants and hands-free copilots. Here is what is in the box, what it costs, and the claims worth checking before you build on it.
What Mistral Voxtral 3 includes
According to Mistral’s Voxtral 3 announcement, the release has three layers: transcription models, a speech model, and an SDK that ties them together with a reasoning model in the middle.
| Component | What it does | Mistral’s stated price |
|---|---|---|
| Voxtral Mini Transcribe 3 (batch) | 4B-parameter speech-to-text for recorded audio | $0.0035/min |
| Voxtral Mini Transcribe Realtime 3 | Streaming transcription with end-of-turn detection | $0.008/min |
| Voxtral Small Transcribe 3 | Higher-accuracy transcription with diarization and context biasing | $0.004/min |
| Voxtral Speech 3 (Mini TTS 3) | Expressive text-to-speech, batch and realtime | $13 per 1M characters |
| Voice agent API with Mistral Small 4 | Full listen-think-speak pipeline | $0.04/min |
| Voice agent API with another reasoning model | Same pipeline, different LLM | $0.08/min |
The same figures appear on Mistral’s API pricing page, which also lists the older Voxtral TTS at $16 per million characters, so the new speech model is cheaper than its predecessor on paper.
Transcription: 50 languages and speaker labels
Mistral says the transcription family now supports 50 languages, including all 24 official EU languages and “many low-resource Asian and Indic languages.” Three features matter most for real deployments:
- Diarization: the model labels who said what, which is essential for call recordings and meetings.
- Context biasing: you can pass key terms or free text, such as product names or staff names, to improve recall on specialist vocabulary without fine-tuning.
- End-of-turn detection: the realtime model uses acoustic and semantic signals to judge when a caller has finished speaking, so the agent can reply quickly without cutting people off.
Mistral claims Voxtral Small Transcribe 3 is state of the art in multilingual transcription and that it outperforms Whisper, ElevenLabs’ Scribe v2 and Gemini 3.5 Transcribe on its charts. Those are Mistral’s own benchmark results, so treat them as the company’s claim until independent tests appear. The realtime model’s documentation lists configurable latency from 80 ms to 2,400 ms and marks it as a public preview.
Speech: emotion, dialects and Pro Voice
On the output side, Mistral says Voxtral Speech 3 supports more than 35 languages and 50 dialects, with ready-made voice presets in 18 languages including Arabic, Hindi, Mandarin, Korean, Vietnamese and several European languages. You can steer delivery with explicit emotion, pacing and emphasis instructions, or let the model infer tone from the text.
Two features stand out:
- Pronunciation hints: Mistral’s demos show phoneme hints that force unusual names or brand words to be said a specific way, a common failure point for customer-support bots reading out email addresses, flight times and prices.
- Pro Voice: a voice adaptation option that, Mistral says, reproduces a speaker’s character and emotional range across languages from about 30 minutes of audio.
Mistral also says its speech model costs less to run than comparable offerings from OpenAI, Deepgram, ElevenLabs and Google. Again, that is the company’s comparison, based on vendor pricing pages it collected on launch day.
How the voice agent SDK works
The new SDK is where Voxtral 3 becomes more than a pair of models. Mistral’s voice agents documentation describes a pipeline that captures the user’s voice, transcribes it, generates an answer with a model, speaks the answer and plays it back. Users can interrupt mid-sentence, and the agent can search the web, query document libraries, call workspace connectors or run your own functions.
Ways to build
- Server or desktop app: the SDK in Python or Node.js handles the full pipeline.
- Web app: a browser TypeScript SDK, with your server creating the session so the API key never reaches the browser.
- Phone lines: SIP telephony, which Mistral says is available to enterprise customers at launch through their account team.
- Your own pipeline: use the realtime transcription, a chat model and streamed speech separately if you want control over each stage.
Because each stage is a separate model, Mistral says you can swap the LLM, the voice or the language without rewriting the agent. That flexibility is also why the agent API has two prices: $0.04 per minute with Mistral Small 4 doing the reasoning, or $0.08 per minute with another model.
Why the EU hosting angle matters
Mistral is leaning hard on sovereignty. Voxtral 3 runs on Mistral’s API through European, US and global endpoints, or on your own infrastructure if you talk to its sales team. For banks, healthcare administrators, insurers and public bodies in Europe, being able to keep call audio inside a chosen environment, and to decide whether recordings are retained at all, can matter as much as raw accuracy.
That fits the pattern of Mistral’s recent releases. The company has pitched its large models as competitive alternatives to US and Chinese labs, as we covered with Mistral’s latest large model, while also losing some customers to cheaper open models, as with Ecosia’s switch to Chinese open-source models.
A crowded week for voice AI
Voxtral 3 did not land alone. Daily roundups such as My AI Guide’s October 20 recap noted that ElevenLabs released new real-time speech-to-speech models on the same day. Voice is quickly becoming the front end for agentic AI, and vendors are competing on latency, price per minute and how well an agent handles numbers and names.
Should you try it?
If you are building a phone or web voice agent and care about European data handling, Mistral Voxtral 3 is worth a test, especially at $0.04 per minute for the bundled pipeline. Before committing, check three things yourself: transcription accuracy on your own accents and jargon, how the agent handles interruptions, and whether the SIP access you need is available on your plan. If you are new to how these systems plug into business tools, our explainer on what MCP is in AI covers the connector side.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
