Mistral Voxtral 3 Review: A Phone-Ready Voice Agent at $0.04 a Minute

Mistral is done selling you models one at a time. Voxtral 3, announced October 20, 2026 (launch coverage's date; Mistral's own page stamps October 22 — worth flagging), is the French lab's first product that sells a finished job: a voice agent that answers phones, checks bills, and books appointments, billed at $0.04 a minute when its own Mistral Small 4 does the reasoning. Under the hood it's a four-model speech family plus a LiveKit-based SDK — and, unusually for Mistral, the company published per-minute pricing for nearly everything.

I've checked its benchmarks against an independent tester and done the per-minute math on a contact-center deployment. Here's the honest review.

Quick answer

• What it is: Mistral's Voxtral 3 pairs speech-to-text and text-to-speech models with a LiveKit-based voice-agent SDK, so an agent can listen, reason with tools, and speak in one real-time streaming pipeline.
• What it costs: batch transcription from $0.0035/min, realtime $0.008/min, TTS $13 per 1M characters, and the full voice-agent service $0.04–$0.08/min — all from Mistral's own price table.
• The headline spec: transcription in 50+ languages, TTS in 35+ languages and 50+ dialects, and sub-500ms realtime latency.
• The catch: every accuracy claim is Mistral's own. Its chart ties its best model with Microsoft AI Transcribe 2, and an independent tester ranks it slightly behind.
• Who it's for: teams building contact-center and service-desk voice agents who want EU hosting, one bill, and a self-deploy option.

What's in the box

Voxtral 3 ships as five line items, each sold separately:

Voxtral 1 launched in July 2025, Mistral's TTS model in March 2026 — Voxtral 3 is the first release fusing both directions of speech with an agent runtime.

The pricing fine print: $0.04 a minute buys a pipeline, not a model

The single most important sentence in Mistral's price table is what the $0.04/min and $0.08/min voice-agent rates include: the whole conversational workflow — transcription, the reasoning model, and speech output. That makes the math legible: 10,000 minutes of agent talk time costs $400 with Mistral Small 4, or $800 with a larger reasoning model (per Search Engine Watch's calculation from Mistral's US-dollar rates).

The asterisks, because there are always asterisks:

• The $0.04 rate is contingent on using Mistral's own LLM. Bring your own reasoning model and the price doubles to $0.08/min. That's a pricing architecture that nudges you toward the full Mistral stack.
• Telephony is separate. SIP calls, carrier charges, and whatever implementation work your team does are not in the per-minute rate (per Search Engine Watch's breakdown). The $400-per-10,000-minutes figure is the AI bill, not the phone bill.
• TTS is billed in characters, not minutes. $13 per 1M characters is cheaper than the older Voxtral TTS's $16 per million (reported by Teckhustlers from Mistral's pricing page) — but character billing punishes chatty agents. A verbose agent that narrates everything it does will drift well above the headline rate.
• Self-deploy has no published price. "To run the full stack on-premises or in a private cloud, talk to our team" is the literal text. Enterprise buyers should assume a sales conversation, not a price list.
• Rivals price one layer cheaper. Artificial Analysis's per-1,000-minute figures put Microsoft AI Transcribe 2 at $1.67 against Voxtral Small's $4.00 — though Microsoft's $0.10/hour is labeled a limited-time offer, so the gap may not survive.

The benchmark check: Mistral's chart, read carefully

On FLEURS 45 (Mistral's chart, vendor-run), macro-average word error rates with language forced versus auto-detected: Voxtral Small Transcribe 3 at 4.0% / 4.1% — which is a tie, not a win, with Microsoft AI Transcribe 2 at 4.0% / 4.1%. Behind them: Voxtral Mini Transcribe 3 (4.5% / 4.6%), Gemini 3.5 Transcribe (4.5% / 6.1%), ElevenLabs Scribe v2 (4.9% / 4.9%), GPT Transcribe (5.6% / 6.3%), and Whisper large-v3 (14.9% / 16.0%).

That "state-of-the-art" framing deserves an asterisk: Mistral's own numbers tie its flagship with Microsoft's model. And an independent tester tells a slightly worse story: Artificial Analysis puts Voxtral Small at 2.1% AA-WER against Microsoft AI Transcribe 2 at 2.0% and Scribe v2 at 2.2%. On speed, Artificial Analysis measures Voxtral Small at 65.8x realtime, far behind Microsoft's 307.5x and level with Scribe's 65.4x.

The diarization numbers are stronger: Mistral reports a cpWER of 18.4% for Small and 18.6% for Mini, against 30.9% for the previous Voxtral Mini Transcribe 2 and 38.4% for Gemini 3.5 Transcribe — a genuine jump. But the realtime model is where Mistral's own chart admits a weakness: at the 480ms setting on a 35-language FLEURS test, Voxtral Realtime scores 6.4% WER against Gemini 3.5 Realtime Transcribe's 4.2% (per AICoder's reading of the announcement). Mistral sells sub-500ms latency; Google sells the better score at that latency. Both readings come from Mistral's own chart — treat them accordingly.

One more flag from Glonce's close reading of the announcement text: the page is inconsistent in naming (Voxtral Transcribe 3 versus Voxtral Small Transcribe 3) and in the audio needed for voice cloning (30 minutes versus "30+ minutes"). Sloppy docs don't change model quality, but they're worth noting from a company asking enterprises to trust it with phone calls.

Pro Voice: cloning that needs 30 minutes of you

Voxtral Speech 3 covers 35+ languages and 50+ dialects, with emotional adaptation, instruction following (prompts can set emotion, pacing, and emphasis), and pronunciation control. The headline feature is Pro Voice: Mistral says it reproduces a speaker's distinctive character, emotional range, and speaking style across languages from 30 minutes of recordings — though no independent review of clone quality exists yet, and the announcement's own wording is inconsistent (30 vs 30+ minutes).

It prices voice cloning as an enterprise feature rather than a consumer toy.

The SDK: LiveKit, SIP, and the sub-second promise

The voice agent SDK is built on LiveKit and connects SIP telephony and WebRTC browser/app sessions in one streaming pipeline: Voxtral Mini Transcribe Realtime 3 listens and detects turn boundaries (acoustic plus semantic end-of-turn prediction, which Mistral says limits false cut-offs when a caller is interrupted), a general model reasons and calls tools, Voxtral Speech 3 replies. No stage waits for the previous one to finish.

The modularity pitch — "each stage is a separate model, so you can change the LLM, the voice, or the language without rewriting the agent" (Mistral's own quote, via Glonce) — is undercut slightly by the pricing, which doubles the per-minute rate if you swap out Mistral's LLM.

Voice is the missing layer of the agentic wave: Walmart's shopping agents and Mastercard's machine-payment rails handle the acting — Voxtral 3 handles the talking.

Mistral's sub-second-turns claim arrives with no conditions, harness details, or test methodology. Interruptibility is the make-or-break UX property of phone agents, so this is the claim I'd verify first in a trial.

EU hosting and self-deploy: the enterprise argument

Voxtral 3 runs on Mistral's European, US, and global API endpoints, and Mistral explicitly offers the full stack on customer infrastructure or private cloud — with a sales conversation, not a price list, as the entry point. OpenRouter already lists the models with EU-hosted, zero-data-retention endpoints (per MarkTechPost's coverage). Organizations can control where conversations are processed and whether audio is retained.

This is the real differentiator: the self-hostable angle matters most to contact centers, health systems, and finance teams that can't route call audio through a shared US cloud. For everyone else, the value proposition is one vendor, one bill, one pipeline instead of wiring transcription, LLM, and TTS vendors together.

What the announcement does not include: anyDon't assume this family ships as open weights.

Voxtral 3 vs the field

Prices per Mistral's announcement; benchmarks are Mistral's own chart except the Artificial Analysis column. Read the per-1,000-minute column as the honest cost comparison: Microsoft undercuts Mistral 2.4x on transcription alone — but only on a limited-time rate.

Pros and cons

Pros: full pricing published in the announcement — rare and welcome; one API replaces a three-vendor integration; EU hosting and a self-deploy path for regulated industries; modular stages; strong diarization numbers; character-billed TTS cheaper than its predecessor.

Cons: every accuracy claim is Mistral's own, and its own chart ties — not beats — Microsoft; independent testing ranks it slightly behind; the realtime model trails Gemini 3.5 Realtime on Mistral's own numbers; the $0.04 rate requires Mistral's LLM; self-deploy pricing is a sales call; no open weights or license linked; sloppy naming inconsistencies in the launch docs.

FAQ

How much does Mistral Voxtral 3 cost?

Per Mistral's price table: batch transcription $0.0035–$0.0040/min, realtime transcription $0.008/min, TTS $13 per 1M characters, and the full voice-agent service $0.04/min with Mistral's own reasoning model or $0.08/min with another LLM. Agent rates cover the workflow, not telephony.

Is Voxtral 3 open source?

The announcement links no downloadable weights and no open license — treat it as an API product until Mistral says otherwise.

How does Voxtral 3 compare to ElevenLabs?

On Mistral's own FLEURS 45 chart, Voxtral Small (4.0–4.1% WER) beats Scribe v2 (4.9%); independently, Artificial Analysis scores them 2.1% vs 2.2% — close. ElevenLabs' Scribe v2 is $0.22/hour ($3.67 per 1,000 min) against Mistral's $4.00. The real difference is packaging: Mistral sells one voice-agent API; ElevenLabs sells separate stacks.

Can Voxtral 3 clone a voice?

Mistral says yes: basic cloning from 30 or more minutes of audio, and Pro Voice — which adapts a speaker's character, emotional range, and style across languages — from 30 minutes of recordings. The announcement's wording is inconsistent (30 vs 30+ minutes), and no independent review of clone quality exists yet.

Can you run Voxtral 3 on your own servers?

Mistral offers self-deployment and private-cloud deployment, but there are no published terms or prices — you talk to the sales team. The EU-hosted API with configurable audio retention is the compliance-friendly option for teams that can't self-deploy.

Key takeaways

• Mistral's Voxtral 3 is a voice-agent product, not a model drop: four speech models plus a LiveKit-based SDK, priced from $0.04/min for the full pipeline.
• The price table is the honest headline — Mistral published per-minute rates for every component, which is rarer than it should be in this industry.
• The benchmarks need an asterisk: Mistral's own chart ties Voxtral Small with Microsoft AI Transcribe 2, and independent testing puts Microsoft slightly ahead.
• The $0.04/min rate requires Mistral's own LLM, telephony is extra, and self-deploy pricing is a sales conversation — read the bundle price as the start of the math, not the end.
• Best fit: contact centers and regulated industries that need EU hosting or self-deploy; everyone else should price the transcription against Microsoft and ElevenLabs first.

Three days after launch, Voxtral 3 reads like a company that finally understands what its enterprise buyers buy: not a WER number, but a phone line that answers. The pricing transparency is genuinely refreshing; the benchmark marketing, less so. Run the trial on your own accents and jargon — Mistral's chart won't do that for you. Published October 23, 2026.