OCR Got 5x Cheaper in 18 Months: A Million Pages Now Costs $65 to $130
Digitizing an archive of paper books and documents now costs pocket change. Daniel van Strien of Hugging Face compared eight open OCR models on 850 pages. The oldest, RolmOCR (April 2025), cost 51 cents per 1000 pages. The newest, LightOnOCR-3-0.8B (October 2026), costs less than 10 cents. A million pages comes to $65–130 in GPU rental. Before you buy a ready-made service, it's worth pricing out your own setup.
AI-processed from The Decoder; edited by Hamidun News
If you have a warehouse full of scans (old contracts, archive cards, digitized books), the question of how much it costs to turn them into text stops being rhetorical. A recent comparison by Daniel van Strien of Hugging Face shows that text recognition in images has become roughly five times cheaper in a year and a half. A million pages now costs $65–130 in GPU rental. That is no longer a project budget; it is a team-dinner line item.
How the test was run
Van Strien took eight freely available OCR models and ran them all on the same set of 850 pages. The set included old natural history books, a 1771 Encyclopaedia Britannica, and both typewritten and printed archival documents. This is material where recognition traditionally stumbles: faded print, unusual typefaces, complex layouts.
All models ran on the same Nvidia RTX PRO 6000 GPU, which the author rented through Hugging Face at $2.75 per hour. That keeps the comparison fair: the price per page is calculated directly from GPU time, not from the rate cards of different clouds.
Here are the results. The oldest model in the comparison, RolmOCR from April 2025, cost about 51 cents per 1000 pages at the same image size. The newest, LightOnOCR-3-0.8B from October 2026, comes in at just under 10 cents. The five models in between barely differ from one another in price. Per million pages, that works out to $65 to $130, depending on how difficult the source material is.
The newer model also read book pages at least as well as the older one. So the lower price was not bought with lower quality, at least on this dataset.
Where the savings really came from
The most practical finding of the test is that switching models is not the only lever. According to van Strien, optimizing the setup itself produced savings of about the same size, and possibly larger. The author means four things:
- feeding pages to the model more evenly, so the GPU doesn't sit idle;
- using smaller scans that the model can still read;
- updating the software the model runs on;
- using a faster GPU.
An everyday analogy is a print shop with a good printer. If you feed in paper in fits and starts and scan images at a resolution nobody needs, even the best printer will run slowly and expensively. So if you already run OCR, first check whether your pipeline sits idle half the time, and only then chase a new checkpoint.
What remains unclear
The result comes with caveats worth keeping in mind. First, the numbers are tied to a specific GPU rental at $2.75 per hour and to a specific set of documents. On your invoices, handwritten manuscripts, or tables with small print, price and quality may differ.
Second, some models in the test behaved badly. They "saw" text on blank pages, meaning they made it up, or they repeated the same fragment endlessly. For an archive, this is a nasty error: it looks like normal text and slips into the database easily. If you are building search or RAG on top of digitized documents, you should catch blank pages and looped repeats with a separate check.
Third, this is one author's comparison, not an independent benchmark. But the direction is clear: open recognition models have become cheap enough to run on your own hardware or on a rented GPU.
What to do in practice
Product teams and founders who pay for cloud OCR by the page should price out the alternative now. Take your own sample of a few hundred pages, run two or three open models on it, and compare cost and quality. Developers should start by tuning the pipeline (scan size, batching, software versions) before swapping the model. In any case, you need a check for blank pages and repeats: at under $130 per million pages, the most expensive problem is not compute but silent errors in your data.
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Open DaisyRose-27B solves 12% of Terminal-Bench tasks, up from 5.6%
- Nov 19, 2026Reson8 Open-Sources a Speech Model 60 Times Smaller Than NVIDIA's: 7.86% Word Error Rate vs. 7.78%
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
Key Questions Answered
How the test was run?
Van Strien took eight freely available OCR models and ran them all on the same set of 850 pages. The set included old natural history books, a 1771 Encyclopaedia Britannica, and both typewritten and printed archival documents. This is material where recognition traditionally stumbles: faded print, unusual typefaces, complex layouts. All models ran on the same Nvidia RTX PRO 6000 GPU, which the author rented through Hugging Face at $2.75 per hour.
Where the savings really came from?
The most practical finding of the test is that switching models is not the only lever. According to van Strien, optimizing the setup itself produced savings of about the same size, and possibly larger. The author means four things: feeding pages to the model more evenly, so the GPU doesn't sit idle; using smaller scans that the model can still read; updating the software the model runs on; using a faster GPU. An everyday analogy is a print shop with a good printer.
What remains unclear?
The result comes with caveats worth keeping in mind. First, the numbers are tied to a specific GPU rental at $2.75 per hour and to a specific set of documents. On your invoices, handwritten manuscripts, or tables with small print, price and quality may differ. Second, some models in the test behaved badly. They "saw" text on blank pages, meaning they made it up, or they repeated the same fragment endlessly. For an archive, this is a nasty error: it looks like normal text and slips into the database easily.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.