Voxtral 3 Guide: Mistral’s Phone-Answering Voice AI — Models, Prices, and Your First Voice Agent

Voxtral 3 is an AI that hears, understands, and talks — launched October 20, 2026 by Mistral AI (per Mistral’s official announcement, as consistently reported by multiple outlets). Think of it as the voice box and ears for AI phone assistants: it transcribes speech to text, speaks back in 35+ languages, and comes with a developer kit that wires the whole loop — hear, think, look something up, answer — into a single pipeline, so a small business can have an AI receptionist answering real calls.

Why this matters to you

Voxtral 3 turns phone calls into a solvable automation problem. A restaurant can let an AI take takeout orders after hours. A hair salon can let it book appointments while the stylist is busy. The building blocks that used to require four different vendors (transcription, brain, tool calls, voice) now arrive as one Mistral package. This is what it means for you: if you can sell websites to local businesses, you can now sell “an AI that answers your phone” — and the unit economics (cents per minute) are cheap enough to mark up 10–20x.

The four models and their prices

Per Mistral’s official announcement, as consistently reported by multiple outlets (vendor figures, not independently verified):

Model	What it does	Price
Voxtral Mini Transcribe 3	Batch transcription — turn a recorded audio file into text	$0.0035/min
Voxtral Mini Transcribe Realtime 3	Live streaming transcription, under 500ms delay	$0.008/min
Voxtral Small Transcribe 3	Highest-accuracy batch transcription	$0.0040/min
Voxtral Speech 3	Text-to-speech (TTS) — reads text aloud, 35+ languages / 50+ dialects	$13 per 1M characters

The voice agent SDK (Software Development Kit — a ready-made code toolkit) sits on top: you pick a reasoning model for the “thinking” step — Mistral Small 4 at $0.04/min, or a bigger reasoning model at $0.08/min. Phone calls go over SIP (the standard protocol that connects internet voice to real phone numbers); web and app calls go over WebRTC (browser-based live audio). Transcription covers 50+ languages. The API was available on launch day across EU, US, and global endpoints.

The money play: an AI phone receptionist for small businesses

Who buys: restaurants, salons, dental clinics, repair shops — businesses that miss calls during busy hours. What you sell: a voice agent that answers missed calls, takes orders or bookings, and texts the owner a summary. What you charge: $300–$500/month flat, plus the few dollars of API cost per business (a salon getting 20 calls/day at 3 minutes each costs you roughly $3/day — well under the fee). 7-day validation: Pick one friendly local business. Build the agent in days 1–3 using the steps above. Offer a free 30-day trial starting day 4. One signed trial = validated; convert or repeat with the next business.

Launch a phone agent in five steps

Ready to build? Here’s the five-step path to a working phone agent.

1. Sign up and get an API key. Go to console.mistral.ai (La Plateforme), create an account, add a payment method, and create an API key. Save it somewhere safe — it’s shown once.
2. Try transcription with one request. Record a short voice note, then send it to the transcription model to see the text come back. This proves your key works before you build anything.
3. Pick your pipeline pieces. For a live agent: Realtime Transcribe ($0.008/min) to hear, a reasoning model ($0.04/min) to think, Voxtral Speech 3 ($13/M chars) to speak. The SDK assembles them into one pipeline.
4. Write the agent’s instructions. Give it a role and the tools it may use. Copy-paste starter:

   “You are the phone receptionist for Sunny Hair Salon. Be warm and brief. You can check the booking calendar and confirm appointments. If a caller asks for anything you cannot do, take their name and number and promise a callback within one hour.”

5. Connect a phone number and test-call it. The SDK connects via SIP. Call the number yourself, book a test appointment, and fix the instructions where it hesitates or rambles.

Pitfalls and limitations

• No self-serve free tier. You need an account with a payment method before anything works — budget a few dollars for testing.
• Benchmark claims are vendor-claimed. Mistral cites numbers like 4.1% WER (word error rate — the share of transcribed words that are wrong) on FLEURS. Treat these as vendor-claimed, not independently verified; test on your customer’s accent and audio quality before promising accuracy.
• Voice agents still make mistakes. Give the agent an escape hatch (take a message, promise a callback) rather than letting it improvise prices or policies it doesn’t know.
• Phone regulations apply. Robo-calling rules vary by country and state. An inbound receptionist agent is far safer than outbound calling — stay inbound.