Mistral introduced Voxtral 3 on Monday midday: a new family of speech-to-text and text-to-speech models, plus an SDK for voice agents that folds listening, reasoning, tool calling and speaking into one streaming pipeline.
Four things at once, in real time
Mistral lays out the problem plainly. A usable voice agent has to hear correctly across languages, accents, speakers and interruptions, work out what the user actually means, sound natural, and then carry out multi-step actions in connected systems. Each piece has worked for years. Doing all four together with no audible lag is the hard part.
The transcription model, Voxtral Small Transcribe 3, covers more than 50 languages, identifies who is speaking, can be biased toward your own vocabulary with keyphrases, and detects when someone has finished talking—without that, an agent either cuts people off or sits there. On the output side, Voxtral Mini TTS 3 brings new voices that adjust emotion, pace and emphasis to the conversation, across 35+ languages and 50+ dialects.
Pricing and fit
Transcription starts at $0.0035 per minute in batch mode. The realtime variant runs $0.008 per minute, covers 40+ languages and is meant to stay under 500ms latency. Speech output is priced at $13 per million input characters.
Mistral names contact centres, service desks, virtual assistants and hands-free copilots as the targets. I wrote about the first Voxtral TTS model back in March, when it shipped with nine languages and open weights. This time the jump is less about voice quality than about the agent scaffolding around it.
The selling point isn’t in the benchmarks
Voxtral 3 runs through Mistral’s EU-hosted API or on your own infrastructure. For European providers, that is the sentence the whole release comes down to: in a customer-service voice agent, what leaves the building isn’t text, it’s the caller’s voice, and anyone processing recordings has to be able to say where they sit and for how long. On model numbers, Mistral won’t always beat the big US labs. On this question, it has the pitch almost to itself.