By Hamza Faisal

Mistral AI is having a very busy week. A day after launching its Timestral time-series forecasting models, the French AI lab has rolled out Voxtral 3 — a family of speech models and a voice agent software kit that turns spoken conversations into working software agents, no separate tools stitched together after the fact.

The release has three moving parts. Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 — a 4-billion-parameter model — handle speech-to-text, joined by Voxtral Mini Transcribe Realtime 3 for live transcription. Voxtral Speech 3 does text-to-speech. Sitting above them all is a voice agent SDK built on LiveKit that fuses listening, reasoning, tool calling and speaking into a single real-time streaming pipeline, available through Mistral's EU-hosted API or deployable on a company's own infrastructure.

Transcription that beats Whisper — on Mistral's numbers

Mistral's headline claim is benchmark performance. The company says Voxtral Small Transcribe 3 posts a 4.1% word error rate on the FLEURS 45-language benchmark when the language is specified — 4.0% when it is inferred — against 16.0% and 14.9% for Whisper large-v3 on the same test. The launch materials also name Scribe v2 and Gemini 3.5 Transcribe as beaten, though Mistral has not published per-language FLEURS breakouts to go with the headline number.

The Small model covers 50 languages, including all 24 official EU languages plus Asian and Indic languages with fewer resources, and adds speaker diarization, context biasing for domain-specific vocabulary without fine-tuning, and end-of-turn detection so agents know when a caller has finished talking.

Sub-500ms latency, priced to undercut

The realtime transcription variant is built for agents: Mistral says it delivers transcription in under 500 milliseconds, including while identifying speakers and applying supplied vocabulary. Each conversational turn completes in under a second, and the agent stops talking when the caller cuts in — the details that make or break a phone call.

Pricing is aggressive. Batch transcription starts at $0.0035 per minute for the Mini model, $0.004 per minute for Small, and $0.008 per minute for the realtime variant. Voxtral Speech 3 costs $13 per 1 million characters — and Mistral positions it as more cost-effective than OpenAI's TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash, naming the competition directly.

Agents that do real work mid-call

The SDK is where Mistral is aiming for differentiation. Rather than a demo chatbot, the kit is built for agents that reach into external systems during a live conversation — checking an invoice, booking an appointment — with function calling, speaker diarization, voice cloning and SIP telephony access baked in. Voxtral Speech 3 handles 35+ languages and 50+ dialects, adjusts emotion, pacing and emphasis to the conversation, and a Pro Voice feature adapts a speaker's voice and style across languages from 30 minutes of audio.

True to form for the European lab, Mistral is pitching sovereign deployment as a feature: companies can run Voxtral 3 on their own servers or call it through EU-hosted APIs, keeping control over where conversations are processed and whether audio is retained. With ElevenLabs valued at $22 billion on the back of voice-agent demand, the fight for the voice layer is officially a race — and Mistral just entered it with the full stack.

By Hamza Faisal