Mistral released Voxtral 3 on October 20, 2026, and the price is what stopped me: $0.004 a minute for its most accurate transcription model. By my math, a one-hour client call would cost 24 cents to transcribe. I have not used any of these models. My client-call notes run through Notion AI every day, so this is an observer’s read of Mistral’s announcement, plus the checks I would run before any client audio went near it.
In this article
- Voxtral 3 is four models and a voice agent kit
- Voxtral 3 pricing puts a one-hour call under 25 cents
- Diarization is the benchmark that matters for client calls
- The realtime model trails Gemini on Mistral’s own chart
- EU hosting helps, but Voxtral 3 is not a local model yet
- Four checks I would run before any client audio goes in
Voxtral 3 is four models and a voice agent kit
Voxtral 3 is a family, not a single model. Mistral’s Voxtral 3 announcement lists two batch transcription models, one realtime transcription model, a text-to-speech model, and a voice agent SDK built on LiveKit that ties listening, reasoning, tool calls, and speech into one streaming pipeline.
The transcription side covers 50+ languages, including the 24 official EU languages. Voxtral Small Transcribe 3 is the accuracy tier. Voxtral Mini Transcribe 3 is the cheaper batch tier, and Mistral gives its size as 4B parameters. Both support diarization, which labels who said what, and context biasing, which lets you pass key terms so names and jargon come out spelled right without fine-tuning.
The speech side, Voxtral Speech 3, covers 35+ languages and 50+ dialects, with voice presets in 18 languages. Mistral also offers voice cloning from 30+ minutes of audio. For a solo consultant, most of that is contact-center tooling. The part that touches my week is transcription, so the rest of this post stays there.
Voxtral 3 pricing puts a one-hour call under 25 cents
Voxtral 3 is priced per minute of audio, and the rates are low enough to change how I think about transcription costs. Here is the published rate card:
- Voxtral Mini Transcribe 3 (batch): $0.0035 per minute
- Voxtral Small Transcribe 3: $0.0040 per minute
- Voxtral Mini Transcribe Realtime 3: $0.008 per minute
- Voxtral Speech 3: $13 per 1M characters
- Voice agent API: $0.04 per minute with Mistral Small 4, $0.08 with any other reasoning model
The arithmetic is mine, not Mistral’s. One hour on the Small model is $0.24. Twenty hours of client calls in a month, which is a heavy month for a one-person shop, comes to $4.80. Even the realtime model lands under 50 cents an hour.
That price moves the decision away from cost and toward trust. The question is no longer whether transcription is affordable. It is whether the transcript is accurate enough, and private enough, to put in front of a client. My rule that a paid tool must deliver about three times the value of the free option still applies, but here the bill is too small to be the deciding factor.
Diarization is the benchmark that matters for client calls
For client calls, speaker accuracy matters more than raw word accuracy, and this is where Voxtral 3 posts its strongest number. On Mistral’s diarization chart, Voxtral Small Transcribe 3 scores 18.4% concatenated minimum-permutation word error rate (cpWER), against 19.0% for Pyannote, 20.3% for Microsoft AI Transcribe 2, 23.0% for Scribe v2, and 38.4% for Gemini 3.5 Transcribe. Lower is better.
Why care? A summary that credits the client with my suggestion, or puts their objection in my mouth, is worse than no summary. Speaker labels decide whose action item is whose. The benchmark set, which includes AMI meeting recordings, Switchboard phone calls, and AISHELL-4, is closer to a real client call than clean read-aloud speech.
On plain transcription the lead is thinner. On FLEURS across 45 languages, Voxtral Small Transcribe 3 posts a 4.1% word error rate with the language specified and 4.0% with it auto-detected. Microsoft AI Transcribe 2 posts the same 4.1% and 4.0%. Whisper large-v3 sits at 16.0% and 14.9%. So Voxtral 3 ties the best rival on its own chart rather than beating it.
Context biasing is the feature I would test hardest. On Mistral’s keyphrase chart, built on Earnings-22 audio, recall for Voxtral Small Transcribe 3 rises from 81.4% without supplied key terms to 94.6% with them. Client calls are full of product names, competitor names, and internal codenames, so that gap decides whether a transcript needs five corrections or fifty.
A cheap transcript that mislabels who said what costs more to fix than an expensive one that gets it right.
The realtime model trails Gemini on Mistral’s own chart
The Voxtral 3 realtime tier is good, not best, and Mistral’s own chart says so. On FLEURS across 35 languages, Voxtral Mini Realtime 3 at a 480 ms delay scores a 6.4% word error rate. Gemini 3.5 Realtime Transcribe scores 4.2% on the same chart. Raising Voxtral’s delay to 2400 ms brings it to 5.7%, and cutting it to 240 ms pushes it to 7.4%.
I respect a vendor chart that shows a competitor ahead. It also clarifies the use case. Realtime transcription matters for live captions or voice agents that must answer within a second. For notes I read after a call ends, batch is cheaper and, on these numbers, more accurate.
Every figure in this post comes from Mistral’s own evaluations. I would treat them as directional until independent testers publish results on messy, real-world audio like this:
- Calls with crosstalk and interruptions
- Speakers with strong accents who switch languages mid-sentence
- Laptop microphones in rooms that echo
EU hosting helps, but Voxtral 3 is not a local model yet
The Voxtral 3 privacy story is about where the audio goes, not about keeping it on my machine. Mistral serves the models through European, US, and global API endpoints, and its announcement says customers can control whether audio is retained. Running the full stack on-premises or in a private cloud goes through Mistral’s sales team.
That last option is aimed at organizations, not one-person shops. On the open-weights side, the Mistral AI page on Hugging Face lists earlier Voxtral checkpoints, but as of October 21 I could not find anything labeled Voxtral 3. So the Ollama route I use for sensitive drafts does not apply here yet.
This lines up with a stance I have taken before. I like strong, low-cost models, but routing client data is a sovereignty decision first and a price decision second. An EU endpoint is a real point in favor for European clients, and I made a similar point about Mistral Large 4 and client drafts. Client audio is still client data, so it falls under the same rules I keep for local AI and client privacy: confidential material stays off the cloud, whatever the per-minute price.
Four checks I would run before any client audio goes in
If I tested Voxtral 3, I would start with recordings I own, not client calls. My current baseline is Notion AI, which I use daily to summarize meetings and pull action items, and I covered that workflow in my notes on Notion AI for client meetings. Any new tool has to beat that setup on accuracy, not just on price.
- Speaker labels: record a 30-minute mock call with a colleague who agrees to it, then count how many turns land on the wrong speaker.
- Context biasing: feed in a glossary of 20 product and company names, then compare misspellings with and without the list.
- Retention in writing: confirm which endpoint processes the audio, whether anything is stored, and what the data processing agreement says.
- Real monthly cost: multiply my actual call minutes by the rate and compare that with what I already pay for meeting notes.
Only if all four pass would I ask a client, say a B2B SaaS founder with a heavy call schedule, whether they were comfortable with a recorded, transcribed call. Consent comes before convenience.
For me, Voxtral 3 is worth watching for the price alone, and the diarization numbers are what would get it a test. The ties and losses on Mistral’s own charts make the announcement more credible, not less. But a cloud API without public weights does not fit the client work I keep local, and my Notion AI setup already covers the rest. I will report back if a mock-call test changes that.
Sources
- Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages — Mistral AI
- Mistral AI organization page — Hugging Face
AI-assisted research and drafting. Reviewed and published by ToolMint.