Voxtral 3 pricing: what Mistral's new voice models cost
2026-10-21 · 4 min read
Voxtral 3 is Mistral's new family of speech models. Transcription costs $0.0035 to $0.008 a minute, speech generation costs $13 per million characters, and a complete voice agent that listens, thinks and talks back runs $0.04 a minute on Mistral's own API. It launched on October 20, and the prices are low enough to change the math on phone and front-desk automation.
What Mistral announced
@MistralAI introduced Voxtral 3 as the transcribe and speech models "powering Mistral's new voice agents API." The second post in that thread highlights turn detection: the transcriber "detects when someone has finished speaking so agents can respond without delay." Most voice bots wait for a fixed stretch of silence, which is why they cut you off or leave an awkward pause.
Mistral cofounder @GuillaumeLample summed up the release in three parts: transcription models for batch and real-time use, an expressive multilingual text-to-speech model, and a voice agent SDK with tool calling.
The first outside numbers came within the hour. @ArtificialAnlys put Voxtral Small Transcribe 3 at number 3 on its speech-to-text leaderboard and placed the TTS model at number 12 in both of its text-to-speech arenas. So transcription is near the top, and the voice is good without being the best you can buy.
Voxtral 3 pricing at a glance
These are the list prices from Mistral's announcement:
| Model | What it does | Price |
|---|---|---|
| Voxtral Mini Transcribe 3 | Transcribes recorded audio | $0.0035 / min |
| Voxtral Small Transcribe 3 | Most accurate batch transcription, speaker labels | $0.004 / min |
| Voxtral Mini Transcribe Realtime 3 | Live transcription for agents, under 500 ms | $0.008 / min |
| Voxtral TTS 3 | Text to speech, batch or streaming | $13 / 1M characters |
| Voice agent API | Listen, reason, reply, with Mistral Small 4 | $0.04 / min |
| Voice agent API | Same, with a larger model | $0.08 / min |
To put that in owner terms: a five-minute call handled end to end by the agent costs about 20 cents on the cheaper tier, and an hour of recorded meetings costs about 24 cents to transcribe with the Small model. Transcription covers 50+ languages. Speech covers 35+ languages and 50+ dialects.
Is Voxtral 3 open source?
Not in the way the first Voxtral was. Mistral's announcement says you can run the models on your own infrastructure, but for on-premises or private cloud it tells you to contact its team, and it does not list downloadable weights or a license. Developers spotted that on launch day: @pHequals7 quoted the launch post with a two-word reaction: "no OSS." If you need a model you can download and run yourself today, Voxtral 3 is not that, at least not yet.
How to try it
- Sign in at console.mistral.ai. AI Studio has playgrounds for voice agents, transcription, speech generation and voice cloning, so you can hear it before you write any code.
- Create an API key on the same console.
- For transcription in code, install the Python SDK with
pip install "mistralai[realtime]"and call the transcription endpoint withvoxtral-mini-transcribe-3-0. The audio docs list the model IDs and a first example. - For a voice agent, set its instructions, voice, tools and welcome message, then run a session. The SDK handles capture, transcription, the reply and playback, and signed webhooks tell your backend when a call starts, ends or fails.
There is one catch for anyone picturing a phone bot. Per Mistral's voice agent docs, phone calls over SIP are open only to enterprise customers at launch, through an account team. Browser and app agents are open to everyone. And if you want the agent to sound like you, voice cloning needs about 30 minutes of sample audio.
How it stacks up
OpenAI's GPT-Live-1.1 worked on voices and fewer pauses inside ChatGPT, and xAI's Grok Voice Transcribe 2 went after transcription accuracy. Mistral is selling the whole pipeline in one place. Each stage is a separate model, so you can swap the language model or the voice without rebuilding the agent, which helps if you expect to change vendors later. It's also Mistral's second release this month aimed squarely at business work, after its Timestral forecasting models.
My read: for most businesses, the transcription price is what counts. At under a penny a minute for top-three accuracy, there is little reason left to type up voicemails, intake calls or meeting notes by hand. The phone agent is promising, but with SIP limited to enterprise accounts and the voice benchmarks graded by Mistral itself, test it on your real calls before customers hear it.
What this means for a small business
With speech models this cheap and this accurate, the technology is the easy part. The harder part is deciding which calls a machine should handle, what it may promise, and when it passes the caller to a person. If you want to work through that before you spend anything, our free process audit starts there.