Voxtral 3 transcription pricing, per hour of audio

Voxtral Mini Transcribe 3 costs $0.0035 a minute on the Mistral API, which is $0.21 an hour of audio. Small Transcribe 3 costs $0.004 a minute ($0.24 an hour) and Mini Transcribe Realtime 3, the live streaming model, $0.008 a minute ($0.48 an hour), all read on Mistral’s pricing documentation on 22 October 2026. That puts the batch hour at 42% below Whisper, but not at the bottom of the market: xAI still sells a recorded hour for $0.10.

The per-minute rates are simple. What decides your bill is which of the three models you call, whether your audio is live, and how much accuracy each extra cent buys.

What Voxtral 3 transcription costs per minute and per hour

Mistral bills transcription by the minute of audio you send, with no plan tiers and no separate line for speaker labels or custom vocabulary. The three Voxtral 3 models launched on 20 October 2026, and their model cards list all three as Public Preview under Mistral’s Premier label.

The three Transcribe 3 models on Mistral’s grid

The Mistral API pricing documentation carries the three rates in its audio grid, next to the TTS model and the previous generation. The second column converts each one to an hour and to 1,000 hours.

Mini Transcribe 3 is the 4-billion-parameter model, and the Voxtral 3 launch post lists it as the batch option, meaning recorded files rather than a live stream. Small Transcribe 3 is the flagship for accuracy. Realtime 3 is the one built to sit inside a voice agent.

One hour on Voxtral Mini Transcribe 3: 60 minutes x $0.0035 = $0.21. On Small Transcribe 3, $0.24. On Realtime 3, $0.48.

Mini Transcribe 2 is still cheaper, and still listed

The previous model, Voxtral Mini Transcribe 2, still sits on the same grid at $0.003 a minute, or $0.18 an hour. So the new generation is not a price cut: Mini Transcribe 3 costs 17% more per minute than the model it follows.

What the extra $0.03 an hour buys is accuracy. On the Artificial Analysis word error rate index, Mini Transcribe 2 scores 3.6% and Mini Transcribe 3 scores 2.2%, so the new model gets about four in ten fewer words wrong. If anyone reads your transcripts, that is the cheapest upgrade in this article.

What the cached column does not tell you

The grid also has a “Cached input” column, at a tenth of each rate: $0.00035 a minute for Mini Transcribe 3, $0.0004 for Small and $0.0008 for Realtime. None of the three model cards says what makes a minute of audio count as cached.

I would plan at the list rate. Audio you have never sent before cannot be a repeat, so for a normal transcription workload the cached rate is a line on the grid, not a discount you can count on.

Is Small Transcribe 3 worth the extra three cents?

Small Transcribe 3 costs $0.24 an hour against $0.21 for Mini, 14% more. On a thousand hours a month that is $30, which is less than a single hour of someone’s time fixing transcripts.

Small against Mini on accuracy

On the Artificial Analysis speech-to-text leaderboard, Small Transcribe 3 posts a 2.1% word error rate and Mini Transcribe 3 a 2.2%. A tenth of a point is about nine words in an hour of speech at 150 words a minute. On English recordings of decent quality, I would not pay for that gap.

Mistral’s case for Small rests on multilingual audio and diarization, where its launch benchmarks show the wider lead. That is Mistral measuring its own model. Artificial Analysis did rank Small third on its leaderboard on launch day:

Mistral has released new Speech models! Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard, and Voxtral 3 TTS enters the top 12 in both Text to Speech arenas

@ArtificialAnlys

https://x.com/ArtificialAnlys/status/2112587905222709432

My rule: Mini for English and French recordings, Small when your audio mixes languages or has several people talking over each other, and test both on 30 minutes of your own audio before you commit. Running the same 30 minutes through both costs under a quarter.

What both models include at no extra charge

The per-minute rate covers more than the bare transcript. Both batch models list these features on their model cards and in the launch post, with no separate price line on the grid:

• Diarization, which labels who said what.
• Context biasing, where you pass key terms and names to improve recall on jargon.
• Timestamps on the transcription endpoint.
• 50 languages, including the 24 official EU languages.

Mistral’s launch thread puts the same features at the centre of its pitch for Small, with a short demo attached:

Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing. It detects when someone has finished speaking so agents can respond without delay.

@MistralAI

https://x.com/MistralAI/status/2112574618586247583

That matters when you compare. ElevenLabs charges $0.05 an hour extra for keyterm prompting on Scribe v2, so its $0.22 hour becomes $0.27 once you use the feature Mistral bundles in.

What Realtime 3 costs once the audio is live

Realtime 3 costs $0.008 a minute, or $0.48 an hour, 2.3 times the Mini batch rate. Its Voxtral Realtime 3 model card lists configurable latency from 80 ms to 2,400 ms, end-of-turn detection, context biasing and diarization in one WebSocket session.

Every vendor charges more for live audio, but not by the same factor. The chart puts the recorded hour and the live hour side by side for the four meters I would consider for a voice agent.

Mistral sits in the middle. xAI’s streaming hour is cheaper than Mistral’s batch hour, and OpenAI’s live meter costs more than twice Realtime 3.

If you are building the whole agent on Mistral, price the bundle too. Mistral lists a voice agent API at $0.04 a minute with Mistral Small 4 doing the reasoning, and its table does not break that price into parts, so check what it covers before you set it against Realtime 3. The speaking side is covered in my Voxtral 3 TTS pricing breakdown.

Live transcription, per hour: xAI $0.20, ElevenLabs Scribe v2 Realtime $0.39, Voxtral Realtime 3 $0.48, OpenAI gpt-live-transcribe $1.02.

How Voxtral 3 compares with Whisper and the rest

Prices below come from each vendor’s own pricing page on 22 October 2026, except where noted. Error rates come from the Artificial Analysis AA-WER index, read the same day, which blends three test sets of agent calls, parliament speeches and earnings calls. One benchmark is one benchmark, and your audio will move these numbers.

Against Whisper and OpenAI’s own meters

Whisper still costs $0.006 a minute, $0.36 an hour, on its OpenAI model page. Mini Transcribe 3 is 42% cheaper and, on this index, makes roughly half as many errors (2.2% against 4.1%). If you call Whisper out of habit, the full picture of what it costs is in my Whisper API pricing page, and this is the strongest reason yet to move.

OpenAI’s newer gpt-transcribe, at $0.27 an hour and 3.3%, closes part of that gap but still loses to Mini on both counts. gpt-4o-mini-transcribe looks cheaper at $0.18, yet OpenAI bills it by tokens and prints the per-minute figure as an estimate, and it has the worst error rate in the table.

Against Scribe v2, Grok and Microsoft

Voxtral 3 is not the cheapest accurate hour. xAI’s Speech to Text, which runs Grok Voice Transcribe 2.0, costs $0.10 an hour on the xAI pricing page and scores 2.3%, a tenth of a point behind Mini Transcribe 3 for less than half the price. I covered that meter in my Grok Voice Transcribe 2.0 pricing page, and on recorded audio it is still the one to beat.

Two more lines deserve a caveat. Microsoft’s MAI-Transcribe-2 posts the best score in the table at 2.0%, but its $0.10 hour is the price Artificial Analysis publishes rather than a rate on a Microsoft pricing page, and the same goes for Gemini 3.5 Transcribe. Mistral’s model card says Small Transcribe 3 outperforms Microsoft AI Transcribe 2; on this index it does not, by a tenth of a point.

Against ElevenLabs Scribe v2 the match is close: $0.21 against $0.22 for the same 2.2%, and Mistral includes the custom vocabulary that ElevenLabs bills separately. Where Mistral wins clearly is hosting. The API runs from European, US and global endpoints, and self-deployment goes through Mistral’s sales team.

Which transcription model I would use for each job

At the rates above, here is how I would route the work.

• Recorded English audio, lowest cost: xAI Speech to Text at $0.10 an hour, with an error rate within a tenth of a point of Voxtral 3.
• Recorded audio that must stay in the EU, or needs custom vocabulary: Voxtral Mini Transcribe 3 at $0.21 an hour, biasing and diarization included.
• Multilingual meetings with several speakers: Small Transcribe 3 at $0.24 an hour, after a 30-minute test on your own recordings.
• A voice agent you assemble yourself: Realtime 3 at $0.48 an hour if you want Mistral end to end, xAI streaming at $0.20 if you only need the transcript.
• Anything still on Whisper: switch. Mini Transcribe 3 cuts the hour from $0.36 to $0.21 and the error rate by about half.

One warning on all three Voxtral 3 models: they are in Public Preview. Mistral has not said when they reach general availability or whether the price changes then, so keep the model id in configuration rather than in code.

Frequently asked questions

How much does Voxtral 3 cost per hour of audio?

$0.21 an hour on Mini Transcribe 3, $0.24 on Small Transcribe 3 and $0.48 on Mini Transcribe Realtime 3, from per-minute rates of $0.0035, $0.004 and $0.008 on Mistral’s pricing documentation.

Is Voxtral 3 cheaper than Whisper?

Yes. Whisper costs $0.36 an hour, so Mini Transcribe 3 is 42% cheaper, and it scores 2.2% against 4.1% on the Artificial Analysis error rate index.

Does Mistral charge extra for diarization?

Not on its pricing grid. Diarization and context biasing are listed as features of the models, and the grid has one per-minute rate for each model with no add-on lines.

Is the old Voxtral Mini Transcribe 2 still available?

Yes, at $0.003 a minute, $0.18 an hour. It is cheaper than Mini Transcribe 3 but scores 3.6% on the same index, against 2.2% for the new model.