Published: 21 October 2026
Mistral AI launched Voxtral 3 on 20 October 2026: five new speech models and a voice agent SDK that answers and places phone calls in real time. The full stack runs through Mistral’s EU-hosted API or on your own servers. A voice agent costs $0.04 a minute when Mistral Small 4 does the reasoning.
Five speech models and a LiveKit-based SDK make up Voxtral 3
Three models turn speech into text. Voxtral Small Transcribe 3 is the most accurate. It labels who is speaking and accepts context biasing: a list of names and jargon it should recognise, without any extra training. Voxtral Mini Transcribe 3 is a smaller model of 4 billion parameters for batch work. Its realtime version streams text and detects when a caller has finished speaking.
Two models turn text into speech: Voxtral Speech 3, in a batch and a realtime version. It adjusts emotion, pace and emphasis to the conversation and covers more than 35 languages. Ready-made voices exist for 18 of them, including Dutch, German and French. A separate option, Pro Voice, adapts a specific person’s voice from about 30 minutes of audio.
The SDK ties the models together. It is built on LiveKit, an open-source framework for live audio. Listening, reasoning, tool calls and speaking run at the same time instead of one after the other. It answers and places calls over SIP, the standard for internet telephony, and works in browsers and apps over WebRTC. During a call, the agent can check a caller’s identity, look up a bill or book an appointment. Mistral says each turn completes in under one second, but it does not publish how it measured that.
Mistral lists these prices and access options:
- Voice agent API with Mistral Small 4 as the reasoning model: $0.04 per minute
- Voice agent API with any other reasoning model: $0.08 per minute
- Transcription: $0.0035 per minute for Mini, $0.004 for Small, $0.008 for Mini Realtime
- Speech output: $13 per million characters
- Hosting: European, US and global endpoints, or on-premises and private cloud through Mistral’s sales team
- Model IDs: voxtral-small-transcribe-3-0, voxtral-mini-transcribe-3-0, voxtral-mini-transcribe-realtime-3-0 and voxtral-mini-tts-3-0
- Try it: the playgrounds in Mistral Studio
Picking a reasoning model other than Mistral Small 4 doubles the minute price
Each stage of the pipeline is separate, so you can swap the language model or the voice without rebuilding the agent. That freedom has a price. With Mistral’s own Small 4 model, the voice agent API costs $0.04 a minute. With any other model, it costs $0.08. For a team that handles 10,000 call minutes a month, that is $400 or $800, by our calculation.
The price table also leaves gaps. It does not say whether the agent price includes transcription and speech output. It names no price for Pro Voice, for a private deployment or for the phone line itself. Ask for those figures before you compare Voxtral 3 with the voice agent you run today.
Voxtral 3’s lead over Whisper and ElevenLabs rests on Mistral’s own charts
Mistral says Small Transcribe 3 beats OpenAI’s Whisper, ElevenLabs’ Scribe v2 and Google’s Gemini 3.5 Transcribe. It also says Voxtral Speech 3 costs less to run than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Almost all of these comparisons come from Mistral’s own evaluations. Until independent results appear, recordings from your own phone lines are the better test.
Mistral is also less open than in March, when Voxtral TTS came with open weights under a non-commercial licence. For Voxtral 3, Mistral does not say whether it will publish the weights. Running the stack on your own servers goes through its sales team.
Can you run a Voxtral 3 voice agent without sending calls to the US?
Yes. Mistral, a French company, offers Voxtral 3 on a European endpoint next to its US and global ones. You can also deploy the full stack on your own servers or in a private cloud. Then the audio stays inside your own environment, and you decide whether conversations are stored.
That matters for the organisations DataNorth works with. Phone calls carry names, addresses and account details, so where the audio is processed is often the first question a privacy officer asks. Voxtral 3 is a complete voice stack with Dutch, German and French voices that can stay in the EU.
Start with one narrow call type that comes in often, such as rescheduling appointments or checking an order status. Give the transcription model your product and place names. Test it on real recordings with regional accents, and measure the turn time on an actual phone line rather than in the browser playground. Under Article 50 of the EU AI Act, callers must be told they are talking to an AI system. Anyone whose voice you clone with Pro Voice should agree in writing first.
In agent projects, the voice is rarely the hardest part. The work sits in the links to booking or billing systems, and in the moment the agent hands a caller to a person. If you need an on-premises setup, ask Mistral for a price before you plan a pilot around it.
For more information, visit the official announcement of Voxtral 3 on the Mistral AI blog.
