The model is Apache 2.0 and costs nothing. The bill is the GPU. Here is the project's own published figure, the same throughput repriced at AWS on-demand rates, and the effective rate at the volume you actually run.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
Uploading...
olmOCR returns text. Drop a document here to see what structured fields with confidence scores look like on the same page.
olmOCR has no price. It is an Apache 2.0 open-weight 7B vision language model from the Allen Institute for AI, so the whole bill is the GPU it runs on. The project's paper reports converting a million PDF pages for $176, which is $0.176 per 1,000 pages, on rented capacity with the fleet running flat out. Reprice that same throughput at AWS on-demand H100 rates of $6.88 per GPU-hour and it becomes about $450 per million, or $0.45 per 1,000. Both numbers assume the card never sits idle. At 100,000 pages a month on a single H100 instance the effective rate is about $50 per 1,000 pages, which is 33 times what a managed OCR API charges.
That is the whole story of open-model economics in one paragraph. The published figure is a throughput number wearing a price tag: it tells you what a saturated GPU fleet costs per page, not what your workload will cost. The distance between those two things is utilization, and for most business document volumes it is a factor of ten or more.
Two primary sources publish it. The olmOCR paper on arXiv 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 on GitHub words it more conservatively as less than $200 USD per million pages converted. Both describe the same thing: a fine-tuned 7B vision model, batched across GPUs, running continuously on rented capacity.
$176
per million pages
The arXiv paper figure, on rented GPU capacity at full saturation. $0.176 per 1,000 pages.
$450
per million pages
The same derived throughput repriced at AWS on-demand H100 rates. $0.45 per 1,000 pages.
$6,240
per million pages
The GPT-4o comparison the paper drew against, at the prices of the time.
The project publishes no throughput figure, so the AWS number above is derived from the two costs it does publish. At roughly $2.69 per H100-hour, the commodity rented rate the $176 figure is built on, a million pages implies about 65 GPU-hours, which is roughly 15,300 pages per H100-hour. Hold that throughput constant and change only the price of the hardware. AWS charges $6.88 per H100-hour on a p5.4xlarge, so the identical work lands at about $450 per million pages. Treat 15,300 pages per GPU-hour as an inference from published costs, not as a benchmark anyone ran. If throughput matters to your decision, measure it on your own documents for two days. That number will hold and a borrowed one will not.
One AWS p5.4xlarge H100 instance at $6.88 an hour, $5,022.40 a month for 730 hours, against the derived ceiling of about 15,300 pages per GPU-hour. The rate is simply the monthly bill divided by the pages you actually ran.
| Pages a month | GPU utilization | Effective per 1,000 pages | Against the market |
|---|---|---|---|
| 11,169,000 | 100% | $0.45 | The paper figure, repriced at AWS on-demand |
| 5,584,500 | 50% | $0.90 | Still well under every cloud OCR rate |
| 2,792,250 | 25% | $1.80 | Now above the $1.50 managed rate |
| 1,000,000 | 9% | $5.02 | 3.3x a cloud API, for a card you also operate |
| 500,000 | 4.5% | $10.04 | 6.7x |
| 100,000 | 0.9% | $50.22 | 33x |
| 50,000 | 0.4% | $100.45 | 67x the published olmOCR figure |
Worth quoting: at 50,000 pages a month, running olmOCR on a dedicated AWS H100 instance costs about $100 per 1,000 pages, which is 67 times the project's own published figure and 67 times what AWS Textract charges for the same pages. Nothing about the model changed. The GPU was simply idle 99.6 percent of the time.
A smaller card fixes part of this. olmOCR needs 12 GB of GPU RAM, so it fits on a 24 GB L4 at $587.50 a month rather than an H100 at $5,022.40, and that moves break-even against a $1.50 API down from about 3.35 million pages a month to about 391,700. The card has to keep up with the volume, which is the constraint nobody can size for you from a blog post, but the direction is clear: for open-model OCR the cheapest adequate card is almost always the right one. The full per-card arithmetic is on the self hosted OCR cost reference.
The README asks for a recent NVIDIA GPU with at least 12 GB of GPU RAM and 30 GB of free disk, and lists the RTX 4090, L40S, A100 and H100 as tested. AWS on-demand Linux rates in US East (N. Virginia), from Amazon's own pricing feed on July 22, 2026.
| GPU | AWS instance | Per GPU-hour | Per month at 730 hours | Break-even pages a month vs $1.50 per 1,000 |
|---|---|---|---|---|
| NVIDIA L4, 24 GB | g6.xlarge | $0.8048 | $587.50 | 391,700 |
| NVIDIA A10G, 24 GB | g5.xlarge | $1.0060 | $734.38 | 489,600 |
| NVIDIA L40S, 48 GB | g6e.xlarge | $1.8610 | $1,358.53 | 905,700 |
| NVIDIA A100, 40 GB | p4d.24xlarge (8x) | $2.7447 per GPU | $16,028.58 | 10,685,700 |
| NVIDIA H100, 80 GB | p5.4xlarge | $6.8800 | $5,022.40 | 3,348,300 |
The RTX 4090 is on the tested list and is not in this table because AWS does not rent one. If you already own consumer hardware, the marginal cost of running olmOCR on it is electricity, and the entire cost argument on this page stops applying. That is the genuinely cheap path, and it is also the one that does not scale past a single machine.
Cheapest first. Read the third column before the second: two rows can share a price and hand back completely different things.
| Option | Per 1,000 pages | What comes back |
|---|---|---|
| olmOCR, rented GPUs at full saturation | $0.18 | Text and markdown |
| olmOCR, AWS on-demand H100 at full saturation | $0.45 | Text and markdown |
| Azure AI Document Intelligence Read, commitment tier | $0.45 | Text, lines, words, confidence |
| GPT-5.4-nano, OpenAI Batch API | $0.84 | Text, whatever you prompt for |
| Azure AI Content Understanding, Basic | $1.00 | Text |
| AWS Textract Detect Document Text | $1.50 | Text and bounding boxes |
| Azure AI Document Intelligence Read | $1.50 | Text, lines, words, confidence |
| Google Enterprise Document OCR | $1.50 | Text and layout |
| Mistral OCR 4 | $4.00 | Text and markdown |
| olmOCR on one H100 at 1,000,000 pages a month | $5.02 | Text and markdown |
| Azure AI Document Intelligence custom extraction | $30.00 | Your fields, with confidence |
| olmOCR on one H100 at 100,000 pages a month | $50.22 | Text and markdown |
The cross-vendor detail behind the managed rows lives on the OCR pricing per 1,000 pages reference, and the language model rows are worked out token by token on the LLM OCR pricing pillar.
The comparison the paper drew was $176 per million pages against $6,240 per million for GPT-4o, a 35 times gap. That gap was the whole argument for building rather than buying, and it is repeated in nearly every article written about open document models since. It is now badly out of date, and not because olmOCR got worse.
Language model pricing collapsed from the other side. GPT-5.4-nano reads a scanned letter page for roughly $1.67 per 1,000 pages at list, and about $0.84 through the Batch API, which is $840 per million. Against the paper's $176 that is a 4.8 times gap, and against the AWS-repriced $450 it is 1.9 times. A saving under two times does not usually survive contact with one engineer's time, an on-call rotation and an upgrade treadmill. The build case for open OCR still exists, but it now rests on data residency and on genuinely large sustained corpora, not on a price gap that has mostly closed.
Worth quoting: the open-model cost advantage over a general purpose LLM fell from 35 times when olmOCR was published to about 4.8 times on rented GPU capacity, and to about 1.9 times at AWS on-demand rates. Anyone citing the 35 times figure in 2026 is quoting a market that no longer exists.
olmOCR is very good at what it was built for: turning PDF corpora into clean, correctly ordered text for retrieval and training. That is a different job from reading an invoice, and the difference is where the cost comparison stops being apples to apples.
olmOCR returns the page as text and markdown in correct reading order. It does not return an invoice number field, a total field or a date field. Anything structured is a second system you write and maintain on top of the text.
A managed OCR API hands back a confidence value per field so you can route the doubtful ones to a person. Open models return a string. Deciding which of those strings to trust is the piece of work that takes the longest to build.
Every real document pipeline needs somewhere for a human to fix the reads the model got wrong. That interface, its permissions, its audit trail and its export are all yours to build, and none of it appears in a per-page comparison.
The published rate assumes a saturated fleet. Real document volume arrives in bursts, month-end spikes and quiet weeks. The GPU bills continuously through all of it, which is why the effective rate is usually several times the headline.
Where DocuOCR sits on this scale is worth saying plainly: about $14 to $20 per 1,000 pages depending on plan, which is far more than a saturated GPU and more than a raw cloud OCR call. What that covers is the four cards above, already built: classification, field-level extraction with confidence values, validation rules, a review queue for the pages that fail them, and export. If your job is converting a static corpus to text, olmOCR is the better tool and it is free. If your job is getting invoice numbers and totals into an accounting system with something checking the doubtful ones, the comparison is not really about the per-page rate at all. Teams that cannot send documents outside their own network should start with on premise OCR software, where deployment rather than price is the decision.
Upload a real document and look at what comes back: the values, the confidence scores and the ones flagged for review. It takes about a minute and it makes the build-or-buy question concrete.
Self hosted OCR cost
Every GPU, every volume, break-even
On premise OCR software
When documents cannot leave your network
DeepSeek OCR pricing
The other open model, costed out
LLM OCR pricing
Every language model, converted to pages
OCR pricing per 1,000 pages
Every vendor, one rate table
OCR API pricing comparison
The buyer view across the market
From the same family of tools
