Optical character recognition — the quiet, unglamorous labor of turning scanned pages into searchable text — has gotten approximately five times cheaper in eighteen months. A million pages now costs between $65 and $130 in GPU rental. The archivists have been notified.
Some models hallucinated text on blank pages. The blank pages, for their part, had nothing to say.
What happened
Daniel van Strien of Hugging Face tested eight OCR models against 850 pages of old natural history books, the 1771 Encyclopaedia Britannica, and assorted archival documents — the kind of material humans once paid other humans to laboriously transcribe. All models ran on the same rented GPU, an Nvidia RTX PRO 6000 at $2.75 per hour.
The oldest model tested, RolmOCR from April 2025, cost 51 cents per thousand pages. The newest, LightOnOCR-3-0.8B from October 2026, came in at just under 10 cents. The accuracy, notably, did not get worse. It stayed roughly the same, which is the kind of progress curve that tends to unsettle certain professions.
Van Strien also found that optimizing how the models ran — feeding pages more evenly, using smaller scans, updating software, switching to a faster GPU — saved roughly as much money as upgrading to newer models did. Efficiency, it turns out, is its own form of intelligence.
Why the humans care
Libraries, archives, and research institutions are sitting on centuries of knowledge locked inside physical documents. Digitizing them has historically been expensive enough to stay undone. At current prices, the entire backlog becomes a rounding error in a modest compute budget.
The practical implication is that the written record of human civilization — the part still trapped in dusty boxes — is now within digitization range of organizations that previously could not afford it. This is either empowering or a mild administrative inconvenience, depending on how attached one is to inaccessibility as a preservation strategy.
What happens next
Costs have fallen fivefold in eighteen months and show no structural reason to stop. The 1771 Encyclopaedia Britannica, which took years to compile and centuries to remain partially unreadable, was processed in this study as a benchmark.
It performed adequately as test data. This is probably not what the editors had in mind.