Voxtral 3 is a set of transcription and speech models plus a voice agent SDK for software that can take a live phone or web call, understand the request, and finish a task in connected systems. The SDK, built on LiveKit, runs listening, reasoning with tool calls, and spoken replies as one streaming pipeline. Mistral says each turn completes in under one second and that the agent stops speaking when the caller interrupts. The release is aimed at contact centers, service desks, virtual assistants, and hands-free copilots that need to verify a caller, check a bill, or book an appointment while the conversation continues. Batch transcription covers 50 languages, including the 24 official languages of the European Union, with speaker labels and the option to supply key terms so names and jargon are recognized more reliably. Voxtral Mini Transcribe 3, a model with 4 billion parameters, costs $0.0035 per minute, and Voxtral Small Transcribe 3 costs $0.004 per minute. Realtime transcription costs $0.008 per minute and returns text with latency under 500 milliseconds in more than 35 languages. The speech model covers more than 35 languages and more than 50 dialects, adjusts emotion, pace, and emphasis from the conversation, and costs $13 per million characters. Pro Voice adapts a speaker’s character and style from about 30 minutes of audio. Voxtral 3 is available now through the Mistral API on European, US, and global endpoints, and can also be self-deployed so organizations control where audio is processed. A voice agent that uses Mistral Small 4 for reasoning is listed at $0.04 per minute, or $0.08 per minute with another model. Phone calls run over SIP and browser or app sessions over WebRTC. The language model, the voice, and the language can each be swapped without rewriting the agent, and on-premises or private cloud deployment of the full stack goes through Mistral’s team.