Voxtral 3 Review: Features, Pricing, Pros and Cons

Most voice AI launches this year have tried to sell you one magic model that hears, thinks and talks at once. On 20 October 2026, Mistral went the other way. Voxtral 3 is not one model. It is a box of parts: three transcription models, a text-to-speech model, and a software kit that wires them to a separate "brain", so you can swap any piece without rebuilding the rest.

Most launch coverage flattened that into a single-model story and lost the point. This Voxtral 3 review untangles the stack, separates Mistral's own benchmark claims from the independent numbers, and works out what it really costs per minute.

Voice AI is moving fast enough that xAI's Grok Voice 3.0 landed just six days before this. Everything below reflects Mistral's announcement and docs as of 30 October 2026, ten days after launch, so confirm current models and prices on Mistral's Voxtral 3 page before you budget around any number here.

In one line: Voxtral 3 is Mistral's voice AI stack, with transcription from $0.0035 per minute, speech at $13 per million characters, and voice agents from $0.04 per minute, hosted in the EU or on your own servers. Worth it? For developers building voice agents who care about cost, data residency or swapping parts, it belongs on the shortlist. For anyone wanting open weights, a voiceover app, or production proof, not yet.

Voxtral 3 at a Glance

The Voxtral 3 Family, Model by Model

Prices are from Mistral's launch page as of 30 October 2026. The docs shorten the speech model's name to "Voxtral TTS 3".

What Voxtral 3 Is, and Why It Exists

Mistral built this stack one stage at a time over about fifteen months, and Voxtral 3 connects them.

• July 2025: the original Voxtral, Small (24B) and Mini (3B), with open weights under Apache 2.0.

• February 2026: Voxtral Transcribe 2, a $0.003/min batch transcriber plus Voxtral Realtime, open weights under Apache 2.0, at $0.006/min.

• March 2026: Voxtral TTS, a 4B speech model in nine languages, with non-commercial weights and an API price of $16 per million characters.

• October 2026: Voxtral 3, every stage upgraded, plus the SDK that joins them.

Why bother? Transcription and speech synthesis were already cheap commodities. The hard part is the agent: listening on a phone line, working out what the caller wants, looking something up, and replying without dead air or talking over anyone. Mistral's own announcement admits end-to-end voice agents remain "difficult".

Then there is the sovereignty pitch. Mistral says the family "is designed for sovereign deployment", through its EU-hosted API or your own infrastructure. For a bank in Frankfurt or a hospital in Lyon, call audio is full of account numbers and diagnoses, and "which country does this recording travel to?" can kill a procurement process before the demo.

It goes beyond server location. For some European buyers, the vendor's nationality matters as much as the data centre's, because US law can reach data that US companies hold abroad. A French vendor offering an EU-hosted API and self-deployment is a genuinely different answer than a US API with an EU region bolted on.

How the Voxtral 3 Pipeline Works (the Part Most Coverage Skips)

Voxtral 3's voice agent is three separate models handing a conversation along, all streaming at once: one listens, one thinks, one speaks.

Think of a component hi-fi system versus a boombox. A boombox is one sealed unit; if the speakers disappoint, you replace the lot. A component system has a separate turntable, amplifier and speakers, so you upgrade the speakers without touching the amp. Grok Voice 3.0 and OpenAI's realtime models are boomboxes. Voxtral 3 is the component system.

Stage 1: Listening, with Voxtral Mini Transcribe Realtime 3

The realtime transcriber turns speech into text as the caller talks. Its harder job is end-of-turn detection, working out when the caller has actually finished, using what Mistral calls "both acoustic and semantic information": how you sound and what you said.

A pause is not always an ending. "My account number is... hang on" is a pause, not a finished thought. Jump in too early and the agent interrupts; wait too long and it leaves dead air. Mistral's docs let developers tune the streaming delay per session, from 80 milliseconds to 1.2 seconds, plus a 2.4-second option.

Stage 2: Thinking, with a separate language model

The transcript goes to a general-purpose language model, in Mistral's example Mistral Small 4, its low-cost text model from March 2026. This stage reasons and acts: function calling, web search, and retrieval over your own systems, so the agent can check an invoice or book an appointment mid-call.

It is the cheap-and-fast versus heavyweight-reasoning trade-off we unpacked in our Claude Fable 5.1 and Claude Opus 5 reviews. Mistral's hosted agent offers Small 4 at $0.04 per minute or an unnamed "larger model" at $0.08. Whether it accepts a non-Mistral brain isn't spelled out, so ask before planning around it.

Stage 3: Speaking, with Voxtral Mini TTS 3

The reply goes to the speech model, which Mistral says matches its tone to the conversation so far. Developers get emotion, pacing and emphasis controls (34 emotion tags and 2 pace tags, per the docs), preset voices in 18 languages, zero-shot cloning, and a "Pro Voice" option built from about 30 minutes of a speaker's audio.

Why "no stage waits" matters

In a naive pipeline, each stage finishes before the next starts, and the delays stack into a robotic pause. Mistral's SDK streams instead, so the voice starts speaking as soon as the first words of the reply exist. Mistral says "each turn completes in under one second", but it has published no measured end-to-end figure, and nobody outside Mistral has tested it.

What swapping parts actually buys you

• Change the brain, keep the voice. Move up to the larger model without re-tuning anything else.

• Change the voice, keep the logic. A new brand voice or cloned voice slots in without rewriting the workflow.

• Change the language. Mistral says this needs no agent rewrite, so one booking flow can serve French and Spanish lines.

• Read the middle. Every turn passes through text, giving you a record of what the caller said and what the agent decided. For compliance and debugging, that is gold.

The trade-off: a speech-to-speech model hears tone directly and can sound more natural, while a pipeline only knows what reached the transcript. It comes down to control versus charm.

What's Actually New in Voxtral 3

Against Mistral's own February and March releases, four things changed:

• Accuracy jumped, independently confirmed. On Artificial Analysis' index, word error rate fell from 3.6% (Mini Transcribe 2) to 2.2% (Mini Transcribe 3), roughly 40% fewer errors.

• Languages widened. File transcription now covers 50+ languages, including, Mistral says, all 24 official EU languages, up from 13. Speech went from nine languages to 35+, plus 50+ dialects.

• The agent SDK is new. Built on LiveKit, it connects to phone lines via SIP, to apps and browsers via WebRTC, and to existing IVR and contact-centre software.

• The weights closed. February's realtime model was Apache 2.0. Voxtral 3 has no downloadable weights, and self-hosting starts with a sales call.

What the Voxtral 3 Benchmarks Actually Say

Short version: transcription accuracy has independent backing, voice quality sits mid-pack on an independent arena, and the full voice agent, the actual headline, has not been independently tested at all.

Mistral's own numbers, read carefully

The FLEURS headline looks huge until you check the comparators. Beating Whisper large-v3 by four times is beating a 2023 model. The comparator that matters is Microsoft's, and with the language given, that is a tie.

Diarization, labelling who said what, is more interesting: a lower cpWER means fewer words pinned on the wrong speaker. Mistral used public datasets, so it is checkable, but it is still Mistral grading Mistral.

The most useful number is keyphrase recall. Context biasing means handing the model a list of names and jargon first, like giving a stenographer a glossary before a board meeting. Mistral's chart shows it lifting key-term recall by about 13 points. If your calls are full of product names, test this first.

Take the speech-quality charts with the most salt. Most are labelled "internal", and all were scored by another AI model (gemini-3.8-flash), not human listeners; only Voice Code Bench came from an outside firm, BeSimple AI. Directional, not proof.

What independent testing shows so far

Independent firm Artificial Analysis has already tested the transcribers, and the result is good: Voxtral Small Transcribe 3 ranks third on its WER Index at 2.1%, ahead of ElevenLabs' Scribe v2 at 2.2%.

The same table shows the catch. Microsoft's MAI-Transcribe-2 is slightly more accurate (2.0%) and far cheaper, $1.67 per 1,000 minutes against Voxtral Small's $4.00, and xAI's Grok Voice Transcribe 2.0 scores 2.3% for the same $1.67. Deepgram's Nova-3 is less accurate but transcribes about 620 times faster than real time, against Voxtral's 63 to 77. Voxtral 3 is top-tier accurate, not the price-performance champion.

On voice quality, Voxtral 3 TTS sits around 15th on Artificial Analysis' arena, well up from March's Voxtral TTS at roughly 48th but behind ElevenLabs. Mistral's claim that it "costs less to run" than Gemini Flash holds against Gemini 3.8 Flash TTS at $16.50 per million characters, but Google's Flash-Lite TTS costs $11 and ranks higher.

The full agent? We found no independent test. Artificial Analysis' Speech-to-Speech Index, topped by Grok Voice 3.0 at 86.9, lists no Mistral entry that we could find as of 30 October 2026. A ten-day-old agent stack has not yet met an independent evaluator or a busy call centre.

What Voxtral 3 Costs

As of 30 October 2026, Voxtral 3 costs $0.0035 to $0.0040 per minute for file transcription, $0.008 for live transcription, $13 per million characters for speech, and $0.04 to $0.08 per minute for voice agents.

Some maths: 1,000 hours of recorded calls a month (60,000 minutes) costs about $210 on Mini, $240 on Small, or $480 transcribed live. An agent handling 10,000 call minutes a month costs $400 on Small 4, or $800 on the larger model. A million characters of speech, very roughly 18 hours of audio by our estimate, costs $13.

Is the bigger sibling worth it?

For most readers, the real rival to each tier is the cheaper tier below it. For file transcription, start with Mini: on the independent index, Small buys 0.1 points of accuracy for 14% more money. Switch only if your own files show a difference.

The agent's $0.04 versus $0.08 choice doubles your bill, and Mistral hasn't named the "larger model". Use Small 4 for scripted flows like bookings and order status, and test the bigger tier only on genuinely messy calls.

One caution: Mistral's launch page doesn't say what the $0.04 per minute includes, such as phone carrier charges or heavy tool use. Its general pricing page mentions a 50% batch discount but lists no Voxtral 3 rates, so ask before assuming it applies.

Where Voxtral 3 Is Strong, and Where It Falls Short

Where it's strong:

• The swappable pipeline, with a text record of every turn.

• Agent price. $0.04 a minute against $0.08 for Grok Voice 3.0 and ElevenLabs' agents, and $0.075 for Deepgram's standard Voice Agent.

• Independently backed transcription accuracy, third on Artificial Analysis' index.

• EU hosting and self-deployment, a real edge for regulated European businesses.

• Phone-ready plumbing: SIP, WebRTC and contact-centre integration.

Where it falls short:

• No open weights, a reversal from February's Apache 2.0 realtime model, and transcription prices rose over Voxtral 2.

• The agent is unproven, with no independent latency or task-success numbers.

• Not the cheapest accurate transcriber. MAI-Transcribe-2 beats it on accuracy and price.

• Mid-pack voice quality, around 15th on the independent arena.

• Privacy takes paperwork. Zero data retention is request-and-approve, and the docs list "Agents" among products it doesn't cover.

• Developers only. No studio, no app, just APIs and an SDK.

How to Actually Get Voxtral 3

As of 30 October 2026:

• Mistral API, via European, US and global endpoints, with Mistral Studio playgrounds for testing.

• Self-deployment, on-premises or in a private cloud, starts with "talk to our team", not a download.

• Not on Hugging Face. The newest Voxtral there is an Arabic version of February's realtime model.

• Not on OpenRouter, which only lists the 2025 Voxtral Small.

• Privacy controls. Mistral says you control whether audio is retained, and zero data retention is available on paid plans by request.

Voxtral 3 vs ElevenLabs, Grok Voice 3.0 and Deepgram

Voice AI is crowded, and Voxtral 3 doesn't win everywhere.

Prices from vendor pricing pages or launch coverage, accuracy from Artificial Analysis, all checked 30 October 2026.

Notice the Deepgram row: swappable pipelines are not unique to Mistral. What is distinctive is the combination, one vendor owning every stage including the brain, at $0.04 a minute, with an EU-first hosting story.

Grok Voice 3.0, launched on 14 October 2026, is the boombox at its best and tops the independent speech-to-speech index. OpenAI's GPT-Live-1, tested by Artificial Analysis in an Astra configuration (our GPT-6 Astra review covers that family), is second at 83.2. If natural, interruptible conversation matters more than control, start there.

ElevenLabs remains the voice quality leader, and if you need voiceovers rather than a phone agent, our ElevenLabs vs Murf AI comparison is the better read. The same goes for voiceover studios like Murf AI, WellSaid, LOVO and Unreal Speech, cloning specialists like Resemble AI and Respeecher, and meeting tools like Krisp. Voxtral 3 is none of those.

Who Should Use Voxtral 3, and Who Should Skip It

Use it if:

• You are building a voice agent for a European business, and "where does the audio go?" keeps coming up in procurement.

• You transcribe lots of multi-speaker audio, like sales calls or meetings, and need speaker labels plus custom vocabulary.

• You want to pick the reasoning model and the voice separately, and upgrade either one later.

Skip it if:

• You need open weights. February's Apache 2.0 Voxtral Realtime is still on Hugging Face.

• You make YouTube narration or ad voiceovers. A studio tool fits better.

• You need production proof today. Give it a quarter.

The Verdict on Voxtral 3

Voxtral 3 is the most sensibly designed voice launch of the month, if not the most dazzling. Mistral didn't build a better boombox. It built the components, priced the assembled agent at half the going rate, and gave European buyers a hosting answer most US rivals can't match as cleanly.

Independent numbers back the transcription story. The voice is merely decent, the weights are gone, and the sub-second agent is still a promise.

Our recommendation: if you build voice agents for EU-facing or regulated businesses, shortlist Voxtral 3 and pilot it on your own calls, starting cheap: Mini Transcribe 3 for files and the $0.04 Small 4 tier for agents. For transcription alone, benchmark it against Microsoft's cheaper option first. And if voice quality is the job, ElevenLabs still owns that room.