OCR for a Million Pages Now Costs $65 to $130
A Hugging Face researcher compared eight free models. A thousand pages cost under 10 cents with the newest model and about 51 cents with the oldest.
Researcher Daniel van Strien from Hugging Face (which hosts open source AI models) has put eight free AI tools through their paces for OCR (optical character recognition), which converts images of text into actual text. These were run on 850 pages of historical natural history books, the 1771 edition of the Encyclopaedia Britannica, and archive material. All were processed using the same graphics processing unit (GPU) — a computer chip designed to run AI algorithms. Running the oldest AI tool, RolmOCR (April 2025), worked out at around 51 cents per 1,000 pages at the same resolution as the latest model. This was LightOnOCR-3-0.8B (October 2026), which came in at just under 10 cents.
So, currently, it will set you back anywhere from $65 to $130 in GPU rental fees to process a million pages of material, depending on what it is. Pricewise, then, cost is no longer the main obstacle for libraries and archives with paper-based collections. As Daniel van Strien points out, optimising the process (by balancing the input of pages, downsizing the scans, keeping software up to date, and using a faster GPU) yielded savings roughly equivalent to, and perhaps greater than, those gained by using more recent versions of the AI tools. The prices quoted here are just for the GPU rental ($2.75 an hour). They don’t include personnel costs, the actual scanning, or quality control. And quality control is important: some of the AI tools hallucinated text where there was none, while others got stuck in loops repeating the same text. The latest AI tool could read printed book text at least as well as the earliest one. Unfortunately, the article doesn’t provide accuracy data for individual models. If you’re thinking about undertaking a major digitisation project, it would be wise to check the output against your own material before trusting it.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles