Mistral's Voxtral 3, launched October 20, 2026, is the first voice stack I've seen where the pricing page answers the architecture question. Transcription from $0.0035/min. Speech at $13 per million characters. A full voice agent pipeline — listen, reason, call tools, speak, streamed as one — at $0.04/min with Mistral Small 4, or $0.08/min with another reasoning model. Built on LiveKit, speaking SIP and WebRTC, so it answers actual phone calls, not just browser demos.
Here's how I'd build on it, and the math I'd check before committing.
Caller (phone / WebRTC) → Voxtral Mini Transcribe Realtime 3 (sub-500ms streaming, $0.008/min) → Reasoning model (Mistral Small 4 in the $0.04 tier) → Your tools: calendar, CRM, order lookup, knowledge base → Voxtral Mini TTS 3 (35+ languages, 50+ dialects, $13/1M chars) → Caller hears a reply, mid-call, while the tool result is still fresh.
The old way was three vendors and three failure points. The Voxtral way is one pipeline with one latency budget. That doesn't remove engineering — it moves it to where it belongs: turn-taking, tool design, and evaluation.
Batch transcription: Mini Transcribe 3 (4B) at $0.0035/min. High-accuracy transcription: Small Transcribe 3 (50 languages) at $0.0040/min. Live transcription: Mini Transcribe Realtime 3 at $0.008/min. Speech output: Mini TTS 3 at $13 per 1M characters. Full agent pipeline: SDK + Small 4 reasoning at $0.04/min. Full agent pipeline with another reasoning model at $0.08/min.
Run your real numbers. At 50,000 agent minutes a month, the $0.04 tier is about $2,000/month before token fees; the $0.08 tier doubles it. At 100,000 batch transcription minutes, Mini costs about $350. If your current vendor charges multiples of that for the same minutes, the pilot pays for itself in the comparison alone.
Diarization. Two voices talking over each other is normal on real calls. If your transcriber can't say who said what, your agent will book the appointment under the wrong name. Voxtral Small Transcribe 3 leads with diarization — test it on overlapping speech, not clean monologues.
Context biasing. Your product names, drug names, and street names are not in any generic vocabulary. Biasing teaches the transcriber your domain terms. Build a bias list from your last 100 support tickets before you judge accuracy.
End-of-turn detection. Knowing when the caller has actually finished speaking is the difference between an agent that listens and one that interrupts. This is the feature users feel in the first ten seconds. Test it with s, filler words, and callers who think out loud.
Don't ship on a benchmark. Ship on your calls: (1) 50 real calls: accents, noise, interruptions, and at least two languages if your customers are multilingual. (2) One real tool wired end to end — a booking, a lookup, a cancellation — and measure resolution rate, not word error rate. (3) Interruption handling: does the agent stop speaking when the caller cuts in, and resume sensibly? (4) Cost at real volume, including the reasoning model, not just the transcription tier. (5) Data residency: can this run on your infrastructure or an EU-hosted API, if your compliance requires it?
A note on honesty: Mistral has not clearly stated which Voxtral 3 weights, if any, are openly licensed. Treat this as a deployable commercial stack with strong self-hosting options, not an open-weights release, until Mistral clarifies. That distinction matters for your procurement paperwork, not just your README.
Keep the speech model swappable. Put telephony, business logic, and tool integrations behind your own orchestration layer, and let Voxtral 3 be one provider behind it. Prices are falling and models improve every few months; the team that can swap a voice provider in a week will outlast the team that married one in a quarter.
Voxtral 3 didn't make voice agents perfect. It made the build-vs-buy math boring — and boring math is what gets projects approved. Run the 50-call pilot, measure resolution instead of accuracy, and let your own audio decide.
By Shivam Kumar, Founder, VisionQuantech. Pricing and feature claims are Mistral's (October 20, 2026 launch); verify on your own audio before migrating production workloads.