How Much Does It Cost to Run an Open Source OCR Model?

Between $0.18 and $0.45 per 1,000 pages if the GPU runs flat out, and between $10 and $100 per 1,000 pages at the volumes most businesses actually process. The model costs nothing. The hardware costs $587.50 to $5,022.40 a month on AWS and bills every hour of it whether a document arrives or not, so the only number that matters is how busy you keep the card.

Almost every published figure for open source OCR describes the first case and gets quoted as if it described the second. That single confusion is responsible for more abandoned self hosting projects than any technical problem, so it is worth walking through the arithmetic properly with real 2026 rates.

What does the GPU cost?

These are AWS on-demand Linux rates in US East (N. Virginia), read from Amazon's own pricing feed on July 22, 2026. Monthly figures assume the instance runs continuously, which is 730 hours.

The last column is the volume at which that monthly bill equals what a $1.50 per 1,000 pages managed API would have charged for the same pages. AWS Textract Detect Document Text, Azure AI Document Intelligence Read and Google Enterprise Document OCR all charge exactly that, which makes it a convenient baseline.

Two things fall out immediately. The expensive cards are the wrong place to start, because an H100 box has to clear 3.35 million pages a month before it saves anything. And VRAM is rarely the constraint: most open document models want 12 GB or less and fit comfortably on a 24 GB card, so the big GPUs buy throughput rather than capability.

What does the model cost?

Nothing. The serious open document models ship under permissive licenses, olmOCR under Apache 2.0 for example, which allows commercial use with no per-page fee and no license key. That is genuinely free and it is also the smallest line in the budget, which is the part that catches people out.

So what does it actually cost per 1,000 pages?

Monthly GPU bill divided by pages actually processed. Here is one g6.xlarge at $587.50 a month across realistic volumes.

Read the top row again. At 10,000 pages a month, the free model on the cheapest AWS GPU costs about 39 times what a paid API charges for identical work. Nothing about the model changed. The card was simply idle almost all month, and idle GPU hours bill at exactly the same rate as busy ones. The full per-card version of this table is on our self hosted OCR cost reference.

Where does the $176 per million pages figure come from?

From the olmOCR paper on arXiv, which states in its abstract that the toolkit converts a million PDF pages for 176 USD against over 6,240 USD per million for GPT-4o. The project README words it more conservatively as under $200 per million pages converted. Both are real, both are primary sources, and both describe a saturated GPU fleet on rented capacity priced well below hyperscaler on-demand rates.

You can reprice it. Those two published numbers imply roughly 65 GPU-hours per million pages at the commodity rate the figure is built on, which is about 15,300 pages per H100-hour. Hold that throughput and change only the hardware price: AWS charges $6.88 per H100-hour, so the identical work costs about $450 per million pages, or $0.45 per 1,000. Still cheap, still assuming the fleet never idles. We show the full derivation on the olmOCR pricing page, including what the same instance costs at 100,000 pages a month, which is about $50 per 1,000.

Is an open model still cheaper than a language model API?

Less than it was, and the change is recent enough that most articles have not caught it. The paper's comparison was $176 per million against $6,240 for GPT-4o, a 35 times gap. Today GPT-5.4-nano reads a scanned letter page for about $1.67 per 1,000 pages at list and roughly $0.84 through the Batch API, which is $840 per million. That makes the gap 4.8 times on rented GPUs and about 1.9 times at AWS on-demand rates.

A saving of under two times, on a workload you now have to operate, is a much harder case than 35 times. The full cross-vendor version of that comparison lives on our LLM OCR pricing pillar.

What is not in the per-page number?

Four things, and together they usually exceed the GPU bill.

• Idle hours. Already covered, and worth repeating because it is the whole story: 730 billed hours a month regardless of volume.
• Retries. Document vision models skip pages, loop on dense tables and occasionally refuse outright. Every retry is GPU time you already paid for, so real throughput sits below any benchmark figure.
• The missing workflow. Open models return text, not fields with confidence scores. Building the extraction, the thresholds, the validation and the review queue on top is the largest hidden cost in the whole exercise.
• Operations. Someone keeps the inference server patched, sizes the batch, and handles the failure at 2am. Something also has to notice within seconds when the endpoint stops answering, because a queue of unprocessed invoices is quiet until it is not.

When does self hosting actually pay?

When volume is above roughly half a million pages a month with a real floor rather than a forecast, when the job is page-to-text conversion rather than field extraction, when documents cannot leave your network, or when you already run GPU inference for something else and the marginal cost is scheduling. Open models are genuinely excellent at converting large static corpora into clean, correctly ordered text, which is what they were built for.

For everything else, and that includes most accounts payable, lending and back-office document work in the United States, the fixed monthly bill is the problem rather than the model. If you are processing tens of thousands of pages a month and need invoice numbers and totals with something flagging the doubtful reads, the comparison is not really about the per-page rate at all. The quickest way to see the difference is to run one of your own documents through a managed extraction and compare the fields and confidence scores against the raw text an open model returns on the same page.