Voxtral vs. GPT-4o-Transcribe: The ASR Price-Performance Trap

On the Artificial Analysis speech-to-text leaderboard, GPT-4o Transcribe sits at a 4.0% word error rate and a list price of $6.00 per 1,000 minutes. Voxtral Mini Transcribe sits at 3.5% WER and $2.00 per 1,000 minutes (Artificial Analysis). One model is three times the price of the other and edges it on accuracy too. You paid triple for a model that is not even the more accurate one.

That single inversion is the whole story of buying speech-to-text at scale. The sticker price per minute is the number every vendor page and every listicle ranks on, and it is the number that misleads buyers most reliably. Accuracy is not a feature you can shop for separately. It is a cost line that hides downstream, in the minutes a human spends fixing transcripts and in the automations that fail silently when a name or a number comes out wrong.

The leaderboard says the quiet part

Start with the raw numbers, because the comparison only gets interesting once you trust them. These are from the live Artificial Analysis ASR board, where speed factors are reported as medians across the last seven days of benchmark trials, so the figures reflect recent measurement rather than a launch-day claim.

Model	WER	Speed factor	Price / 1K min
ElevenLabs Scribe v2	2.2%	33.7x	$3.67
Voxtral Small (Mistral)	2.8%	65.9x	$4.00
AssemblyAI Universal-3 Pro	3.1%	86.9x	$3.50
Voxtral Mini Transcribe	3.5%	71.7x	$2.00
Voxtral Mini Transcribe 2	3.6%	74.9x	$3.00
AssemblyAI Universal	3.8%	113.6x	$2.50
GPT-4o Transcribe (OpenAI)	4.0%	33.1x	$6.00
GPT-4o Mini Transcribe (OpenAI)	4.5%	48.1x	$3.00
Deepgram Nova-3	5.2%	520.0x	$4.30

Read down the price column and the ranking looks one way. Read down the WER column and it scrambles. The two columns disagree, and that disagreement is where money leaks. OpenAI API's GPT-4o Transcribe is the most expensive model in this set and ranks seventh of nine on accuracy. GPT-4o Mini Transcribe, at a mid-tier $3.00 per 1,000 minutes, posts the second-worst WER on the board at 4.5%, behind only Deepgram Nova-3's 5.2%.

Look at the two ends of the price column. The cheapest model is Voxtral Mini Transcribe at $2.00, and it is also near the top on accuracy at 3.5% WER. The most expensive is GPT-4o Transcribe at $6.00, and it lands mid-pack. Price and accuracy simply do not track each other on this board. That is the trap in one sentence.

Why WER is a price, not a spec

Word error rate gets treated as a quality badge, a number you would like to be lower the way you would like a benchmark score to be higher. That framing is wrong for anyone running ASR at volume. WER is a price tag denominated in human attention.

Every error a model makes is a word someone downstream has to catch and correct, or a word that flows uncaught into whatever the transcript feeds. At 4.5% WER, roughly one word in twenty-two is wrong. At 2.8%, it is one in thirty-six. On a single five-minute clip the difference is invisible. Across a million minutes of call recordings, the gap is the difference between a review queue you can staff and one you cannot.

Walk one example through. Take 1,000 minutes of recorded calls at maybe 150 spoken words a minute, so roughly 150,000 words of transcript. At 3.5% WER that is about 5,250 wrong words; at 4.5% it is about 6,750. The raw gap of 1,500 errors looks small against a six-figure word count. The cost is not in the gap. It is in the fact that a reviewer has to read all 150,000 words to find any of them. The error rate sets how many corrections you make, but the volume sets how long the search takes, and the search is the expensive part.

The errors are not evenly distributed across word types either. ASR models miss proper nouns, account numbers, drug names, and domain jargon at far higher rates than common words. Those are exactly the tokens your downstream automation keys on. A model can post a respectable global WER and still fumble a meaningful share of the named entities a routing rule or a billing lookup depends on, so the error rate your operations team actually feels is higher than the headline number suggests.

The effective-cost formula

The fix is to stop comparing list prices and start comparing total cost. The model is simple enough to run on a napkin, which is the point. Anyone evaluating vendors can compute it before signing a contract.

effective_cost_per_1K_min =
    api_price_per_1K_min
  + (correction_minutes_per_1K_min × labor_rate_per_min)

where correction_minutes scale with WER
  and with how much of the output a human must touch

The api_price term is the number on the pricing page. The second term is the one nobody prints, and it is usually larger. A reviewer correcting a transcript does not spend time proportional to the error count alone. They spend time finding the errors, which means reading or relistening to material that was mostly fine. That is why a small WER gap produces a large cost gap: the search cost dominates the fix cost.

The model with the lower sticker price wins the spreadsheet and loses the quarter, because the line nobody printed is the largest one in the budget.

Plug in the leaderboard numbers and the inversion becomes concrete. Voxtral Mini Transcribe undercuts GPT-4o Transcribe on the api_price term by two-thirds, from $6.00 down to $2.00, and on the correction term by virtue of a lower WER. There is no labor rate at which GPT-4o Transcribe catches up on a pure batch-transcription workload, because it is behind on both terms of the sum. On batch work, the premium price buys nothing measurable.

GPT-4o Mini is the more instructive case. Its $3.00 sticker matches Voxtral Mini Transcribe 2 exactly. But its 4.5% WER is the second-worst on the board, against 3.6% for the Voxtral model at the same price, so its correction term is the largest in the set. Two models, identical list price, and the one that looks like a tie on the pricing page is the most expensive option you could choose once a human enters the loop. The pricing page cannot show you that, because the term that separates them lives in your review queue, not on the vendor's site.

What Voxtral actually ships now

The "Voxtral" in the title is now a lineup, not a single model. Mistral AI shipped Voxtral Transcribe 2, also branded Voxtral Mini Transcribe V2, on February 4, 2026 (Mistral). The batch endpoint lists at $0.003 per minute and lands around 4% WER on the FLEURS multilingual benchmark per Mistral's own figures, which tracks with the leaderboard's 3.6%.

The release added the features that separate a transcription toy from a production service: speaker diarization, word-level timestamps, context biasing for up to 100 terms, 13 languages, and recordings up to three hours (Mistral). Context biasing is the underrated one. It is the lever that pulls down the proper-noun miss rate that quietly drives your correction cost, by letting you preload the account names and product SKUs a generic model would butcher. If your audio is full of internal jargon, that single feature can move your effective cost more than a tenth of a point of headline WER.

There is also Voxtral Realtime for streaming at $0.006 per minute, shipped under Apache 2.0, a 4-billion-parameter model on Hugging Face that runs at the edge. Latency is configurable down to sub-200ms, and Mistral reports it stays within 1-2% WER at 480ms latency (Mistral). That open-weight, edge-deployable profile matters for the use-case split below, because it changes the math from per-minute API spend to amortized GPU cost. A closed API meters you forever; an open model you host converts the bill into hardware you already own once utilization is high enough.

How OpenAI prices the same job

OpenAI prices transcription by tokens, not minutes, which makes apples-to-apples comparison harder than it should be. gpt-4o-transcribe runs $2.50 per million audio input tokens and $10.00 per million text output tokens; gpt-4o-mini-transcribe input is $1.25 per million audio tokens (OpenAI). The token-to-minute conversion is not something you can read off the pricing page directly, because it depends on how many audio tokens a minute of speech consumes and how much text the model emits, and those vary with the audio itself.

That is why the per-minute figures everyone quotes come from third-party estimation rather than OpenAI's own table. Independent guide coverage estimates roughly $0.006 per minute for the full model and $0.003 for mini (TokenMix), consistent with the leaderboard's $6.00 and $3.00 per 1,000 minutes once you multiply through. Treat those numbers as a working approximation, not a contract: your real bill turns on your own audio length and verbosity, which is one more reason to test on your own files rather than trust a derived per-minute rate.

OpenAI positions GPT-4o Transcribe as an accuracy and language-recognition improvement over the original Whisper models (OpenAI), which is true and also beside the point. The relevant question is not whether it beats two-year-old Whisper. It is whether it beats the current field at its price, and the leaderboard says it does not.

One real constraint cuts in OpenAI's favor for a specific job: the standard transcription endpoint does not stream, and the guidance for real-time work points you to Deepgram or AssemblyAI instead (TokenMix). That is not an accuracy verdict. It is an architecture one, and it splits the buying decision cleanly between work that can wait and work that cannot.

Segment by use case, not by vendor

The effective-cost formula has a different answer for each workload because the correction term and the latency requirement change. Three segments cover most buyers.

Batch transcription is the pure accuracy play: podcasts, depositions, meeting archives, call-center QA. Latency does not matter; WER dominates the total-cost equation because every error becomes a correction minute. Here the cheapest sticker is irrelevant and the accuracy-per-dollar leaders win. Voxtral Small at 2.8% WER and ElevenLabs Scribe v2 at 2.2% are the frontier; AssemblyAI Universal-3 Pro at 3.1% and $3.50 is the balanced pick. Mistral claims its Voxtral Mini batch endpoint processes audio roughly 3x faster than Scribe v2 at one-fifth the cost while matching quality (Mistral). The leaderboard's Voxtral Small row does not corroborate that specific claim, so verify it against the exact Voxtral SKU and your own audio before you treat it as settled.

Real-time voice agents invert the priority. A lead on accuracy is worthless if the model adds hundreds of milliseconds of latency to a live conversation, because the human on the line hears the lag and starts talking over the agent. Streaming support and tail latency are the gating specs, which is why the batch leaders are often the wrong answer and Deepgram Nova-3, with its 520x speed factor, earns its place despite a 5.2% WER. Voxtral Realtime's sub-200ms configurability puts Mistral in this segment too, with the open-weight option of running it on your own hardware to shave the network hop entirely.

Diarization-heavy work is its own segment because "who said what" is a separate quality axis from word accuracy. A transcript with perfect words and scrambled speaker labels is useless for a sales-call summary or a multi-party legal record, and no global WER number tells you how often the labels are right. Voxtral Transcribe 2's built-in diarization and word-level timestamps matter more here than its WER edge, because they remove a whole class of correction work that a words-only model leaves on your desk.

The API versus finished-app decision

Before comparing two APIs, confirm you want an API at all. The leaderboard is a market for raw transcription engines you wire into your own product. If your actual need is "transcribe my team's meetings and give me searchable notes," a finished application like Otter.ai bundles the ASR, the UI, the storage, and the summarization into one subscription.

The trade is control versus integration cost. An API gives you the lowest per-minute price and full control over the pipeline, at the cost of building everything around it. A finished app costs more per minute on paper and saves you the months of engineering that the per-minute math conveniently ignores. The effective-cost formula has a third hidden term for API buyers: the build. If you are transcribing at hobby volume, the build cost never amortizes and the app wins; if you are running a transcription pipeline as a core product, the API's lower per-minute floor eventually pays back the engineering and the app's bundled markup becomes the thing you are overpaying for.

Deployment path changes the math

For enterprises already routing AI spend through a cloud, the provider matters as much as the model. Azure OpenAI Service is a deployment path for GPT-4o transcription that may carry committed-spend discounts and data-residency guarantees a direct API does not, which can move the real price well off the public list. If your organization already has an Azure commitment, the relevant comparison is your negotiated rate against Voxtral's list, not list against list. The published number is a starting point, not the price you pay at scale.

Voxtral's Apache 2.0 realtime model points the other way. At high enough volume, the per-minute API price stops being the relevant number, because self-hosting a 4B-parameter model on owned GPUs converts a variable cost into a fixed one. The break-even depends on utilization, and below it the cloud API is cheaper because you are not paying for idle hardware. Above it the owned GPU wins and keeps winning. The point is that the option only exists with open weights, which is a strategic edge the closed models cannot match regardless of their sticker price.

Run the test before you sign

Pull a representative hour of your own audio, the messy kind with crosstalk and jargon, not a clean studio sample. Run it through Voxtral Mini Transcribe 2, GPT-4o Transcribe, and one accuracy leader like AssemblyAI Universal-3 Pro. Then have someone correct each output to ground truth and put a stopwatch on the correction pass. That stopwatch is your real labor_rate term, and the elapsed minutes on each model are the only number that decides the contract.