Mistral Voxtral 3 is a cheap, wide-ranging speech stack, but the claims of being best are Mistral’s own. Mistral launched it on 20 October 2026: three transcription models from $0.0035 a minute, a text-to-speech model at $13 per million characters, and a voice agent SDK built on LiveKit. Here are the prices, the languages and what each costs in practice.
Key takeaways
- What launched: Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, Mini Transcribe Realtime 3 and Mini TTS 3, plus a voice agent SDK.
- Voxtral 3 pricing: $0.0035 to $0.008 per minute for transcription, $13 per million characters for speech, and $0.04 to $0.08 per minute for voice agents.
- Languages: 50 for transcription, including Hindi and several Indian languages; 35+ languages for speech. Tamil isn’t on Mistral’s tested list.
- Claims: “state of the art” is Mistral’s wording, based on its own charts.
- Where: Mistral’s API through EU, US and global endpoints. On-premises needs a talk with Mistral’s team.
Last updated: 1 November 2026 · Status: live · I’ll update this post when independent rankings or licence details appear.
[Insert “Table of Contents by Rank Math” block here]
What is Mistral Voxtral 3?
According to Mistral’s launch post, Voxtral 3 is a family of speech-to-text and text-to-speech models plus a voice agent SDK that joins listening, reasoning, tool calling and speaking in one real-time pipeline. It targets contact-centre agents, service desks, virtual assistants and hands-free copilots.
The pitch is control. Mistral says the Voxtral 3 speech models can run through its EU-hosted API or on your own infrastructure, so you decide where audio is processed and whether it’s kept.
Voxtral 3 pricing and the four models
| Model | What it does | Price |
|---|---|---|
| Voxtral Mini Transcribe 3 (batch) | Efficient transcription, 4B parameters, 50+ languages | $0.0035 per minute |
| Voxtral Small Transcribe 3 | Highest accuracy, diarization, context biasing, 50+ languages | $0.0040 per minute |
| Voxtral Mini Transcribe Realtime 3 | Live transcription, under 500 ms latency | $0.008 per minute |
| Voxtral Mini TTS 3 | Expressive speech, 35+ languages, 50+ dialects | $13 per million characters |
| Voice agent API with Mistral Small 4 | Listen, reason, call tools, speak | $0.04 per minute |
| Voice agent API with a larger model | Same pipeline, bigger reasoning model | $0.08 per minute |
For comparison, Mistral’s earlier Voxtral TTS launch in March listed speech from $0.016 per 1,000 characters, which is $16 per million. The new $13 rate is about 19% lower. That comparison is mine, not Mistral’s, and the models are different.
What Voxtral 3 pricing means in practice
- Transcribe one hour of audio with Small Transcribe 3: 60 minutes × $0.004 = $0.24.
- Run 1,000 minutes of voice agent calls on the Mistral Small 4 tier: 1,000 × $0.04 = $40. On the larger-model tier: $80.
- Generate speech for a 1,000-word script, roughly 6,000 characters: about $0.08 at $13 per million.
These are list prices before tax, taxes and any volume deals. Multiply by today’s USD-INR rate for rupees. A real call-centre bill also includes telephony, your own tools and any extra model usage.
The Voxtral voice agent SDK, explained
The Voxtral voice agent SDK is built on LiveKit. Mistral says speech in, reasoning and speech out run as one streaming pipeline, so a turn completes in under a second and the agent stops talking when the caller cuts in. The agent can verify a caller, check a bill or book an appointment through function calling, web search and retrieval from your own systems. It connects by SIP phone lines or WebRTC in browsers and apps.
Each stage is a separate model, so you can swap the language model, the voice or the language without rewriting the agent. That’s a real advantage if you don’t want to be locked to one vendor, though you still carry the integration work.
Languages: does Mistral Voxtral 3 work for Indian languages?
Mistral says transcription covers 50 languages, including many low-resource Asian and Indic languages. Its benchmark list names Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi and Telugu. Tamil doesn’t appear on that list. For speech output, Hindi is among the 18 languages with ready voice presets.
If you work in Tamil, don’t assume. Test a few minutes of your own audio, with your accent and background noise, before you commit.
Mistral Voxtral 3: claims versus independent proof
- Mistral says Small Transcribe 3 outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe, and that its TTS costs less than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.
- Mistral’s charts use public benchmarks and some internal evaluations, with an AI model as judge on the speech tests. That’s useful but not neutral.
- Independent check: look at Artificial Analysis’s speech rankings for the current word-error-rate placing. I haven’t repeated their numbers here because they can change.
- Maturity of the Voxtral 3 speech models: one reviewer reports every Voxtral 3 model is still a public preview. I couldn’t confirm that on Mistral’s page.
Watch: Mistral Voxtral 3, explained
[Paste your YouTube or Instagram video link on an empty line below.]
What we don’t know yet
- Whether the models have downloadable weights and under which licence. Mistral mentions self-deployment and on-premises use through its team.
- Where Voxtral 3 ranks on independent speech leaderboards over time.
- How well it handles Tamil and other languages missing from Mistral’s list.
- Whether any model is still a preview. Mistral’s launch post doesn’t say, though one reviewer does.
FAQ
How much does Mistral Voxtral 3 cost?
Transcription costs $0.0035 to $0.008 per minute, speech costs $13 per million characters, and the voice agent API costs $0.04 to $0.08 per minute.
Does Mistral Voxtral 3 support Hindi?
Yes. Hindi is on Mistral’s transcription language list and has speech voice presets. Mistral lists several other Indian languages for transcription.
Does Voxtral 3 support Tamil?
Tamil isn’t on Mistral’s published benchmark list. Test it on your own audio before you rely on it.
Can I run Mistral Voxtral 3 on my own servers?
Mistral says you can self-deploy or run the full stack on-premises or in a private cloud, by talking to its team. It doesn’t state a public licence.
Hari’s take
Mistral Voxtral 3 is aggressively priced and well-packaged for builders who want a European option. The numbers to trust are the prices. The “state of the art” claims need outside checks, and the Indian-language story needs your own test clips. If you build voice tools, run a one-hour trial against what you use today before you switch.
Which language would you test first? Tell me in the comments.
Sources: Mistral: Voxtral 3, Mistral: Voxtral TTS, eesel, Shattered, Reflekt Lab, OpenRouter. Last updated 1 November 2026.
