Mistral AI introduced Voxtral 3 on Oct. 20 with a $0.04-per-minute voice-agent service that combines transcription, reasoning and spoken replies for developers building live-call assistants.
That rate applies when Mistral Small 4 handles the reasoning. Selecting another reasoning model raises the charge to $0.08 per minute. Both rates cover transcription and the broader conversation workflow, rather than one speech function alone.
The service turns a caller’s speech into text, sends the request to a reasoning model and produces a spoken response. An agent can also use connected tools to look up information or carry out a task during the call. Mistral’s software development kit can link an agent to telephone calls via SIP and to browsers or apps via WebRTC.
Developers can buy the speech functions separately. For batch use, Voxtral Mini Transcribe 3 has a $0.0035-per-minute rate; Voxtral Mini Transcribe Realtime 3 has a $0.008-per-minute rate. Voxtral Small Transcribe 3 is priced at $0.0040 per minute.
Voxtral Speech 3, the speech-generation model, is billed at $13 per million characters. Its character-based charge measures a different service from the voice-agent API’s per-minute rate. Developers choosing individual models would pay for the speech functions they use, while the agent rate covers the combined conversation workflow.
Voxtral began with speech-understanding models in July 2025; Mistral added a text-to-speech model in March 2026. With Voxtral 3, an agent can receive a caller’s speech, generate a spoken reply, use a reasoning model and work with connected tools in the same exchange.
The API is available through U.S., European and global endpoints. Organizations must arrange with Mistral to deploy the full system on in-house infrastructure or a private cloud.
