Mistral AI published Voxtral 3 on its official blog on October 20, 2026: a family of speech models plus a voice agent SDK that puts transcription, reasoning, tool calling and speech synthesis into a single real-time streaming pipeline. Transcription covers more than 50 languages, synthesis supports over 35 languages and 50 dialects, and the whole stack is available today through Mistral's EU-hosted API or for deployment on a customer's own infrastructure.
Four pieces, not one model
Transcription comes in three tiers: Voxtral Mini Transcribe 3 at $0.0035 per minute, the Small version at $0.004, and a realtime variant with sub-500ms latency at $0.008 per minute. In Mistral's published FLEURS test across 45 languages, the Small model posts a 4.1% word error rate with the language specified and 4.0% with automatic detection, which the company says beats Whisper large-v3, Scribe v2 and Gemini 3.5 Transcribe; these are vendor-run numbers awaiting independent replication. On the output side, Voxtral Speech 3 costs $13 per million characters, adjusts emotion, pace and emphasis to context, and offers Pro Voice, which reproduces a speaker's vocal character and style from 30 minutes of audio. The LiveKit-based agent SDK ties the pieces together: during a live call the agent can check a bill, book an appointment or verify identity, each turn completes in under a second, and it stops talking when the caller cuts in.
A same-day collision with Eleven Realtime
The timing is pointed. Earlier the same day, ElevenLabs launched Eleven Realtime for live conversation, covered in our report on ElevenLabs' Eleven Realtime release with its latency and reasoning scores. The two products take different shapes: ElevenLabs sells speech-to-speech models, while Mistral sells an assemblable stack in which transcription, the reasoning model and synthesis can each be swapped, and the whole thing can move into the customer's own data center. For European buyers that second point may matter more than benchmarks: support calls carry heavy personal data, and EU hosting or private deployment makes the compliance conversation much shorter.
Which bill to check first
Three concrete checks for teams evaluating voice stacks. First, language distribution: if your users concentrate in lower-resource languages and dialects, Voxtral's breadth is the differentiator, so test error rates on your own call recordings rather than trusting FLEURS alone. Second, the deployment boundary: whether call audio may leave your premises decides API versus self-hosting, and getting this wrong makes everything after it a compliance cost. Third, the price structure: transcription under half a cent per minute and a voice agent API at $0.04-0.08 per minute sit in a different league from per-seat contact-center pricing at volume, but the reasoning model's cost has to be counted in. Voice-agent competition has moved from sounding human to finishing the job inside the customer's data boundary, and Voxtral 3 is a bet on the second question.
