Mistral Ships Voxtral 3, a Voice Agent Stack That Listens, Reasons and Replies in Under a Second at $0.0035 a Minute

Mistral AI has released Voxtral 3, its second-generation speech family and the most complete version yet: three transcription models, a text-to-speech model and a Voice Agent SDK that wires them into a single realtime pipeline. Transcription coverage jumps from 13 languages in the previous generation to more than 50, including all 24 official EU languages and a set of low-resource Asian and Indic languages. Speech synthesis covers more than 35 languages and over 50 dialects.

The architecture is deliberately modular rather than end-to-end. A transcription model listens and detects when the user has finished speaking using both acoustic and semantic cues; a general-purpose reasoning model such as Mistral Small handles intent and tool calls; and the TTS model speaks the response. The LiveKit-based SDK orchestrates the loop over SIP for telephony and WebRTC for browsers and apps, supports interruptions at any point, and completes each turn in under one second, according to Mistral. Developers can swap the reasoning model or the voice model without rewriting agent logic.

Pricing is the sharpest edge. Batch transcription on the 4-billion-parameter Mini model costs $0.0035 per minute, realtime transcription $0.008 per minute, and the larger Small Transcribe 3 model $0.0040 per minute. Synthesis is priced at $13 per million characters, and the voice agent API runs $0.04 per minute with Mistral Small 4 as the reasoning model, or $0.08 per minute with any other model.

Two features target contact-center work directly. Context biasing lets developers inject keywords or free text at call time so product names and internal jargon are recognized without fine-tuning, and Mistral reports materially higher keyphrase recall on its Earnings 22 benchmark when prompts are supplied. The Mini TTS 3 model adjusts emotion, pacing and emphasis from conversational context and can clone a voice from as little as 30 minutes of reference audio.

Performance claims deserve caution because most come from Mistral's own evaluations. The company says Small Transcribe 3 beats Whisper, ElevenLabs' Scribe v2 and Gemini 3.5 Transcribe on word error rate and speaker diarization in its FLEURS 45 tests, using gemini-3.8-flash as the judge. In independent non-realtime testing by Artificial Analysis, however, Small Transcribe 3 posted a 2.1 percent word error rate that ranks third behind models from Alibaba and Microsoft, and its synthesis ranked 13th. These are vendor self-reported results where noted and have not been independently verified across enterprise workloads.

The sovereignty pitch may matter more than the benchmarks. The stack is available through EU-hosted APIs or for self-deployment on corporate infrastructure and private clouds, letting regulated industries control where audio is processed and whether conversations are retained. That positioning is aimed squarely at GDPR-bound contact centers and public-sector service desks that cannot route customer calls through US clouds.

The release tightens competitive pressure on OpenAI, Deepgram and ElevenLabs at the low end of the voice market while making a strategic argument: voice agents are an infrastructure problem, and whoever owns the full pipeline from microphone to tool call, at self-hostable economics, sets the margin structure for the entire category. For call-center operators watching per-minute costs, the math of automated tier-one support has shifted again.