Daniel van Strien of Hugging Face tested eight open-source optical character recognition models on 850 pages from old natural history books and the 1771 Encyclopaedia Britannica to measure processing expenses. He ran all tests on a single Nvidia RTX PRO 6000 GPU rented for $2.75 an hour, finding that the oldest model from April 2025 cost about 51 cents per 1,000 pages while the newest from October 2026 reached just under 10 cents. This shift means digitising a million pages now costs between $65 and $130 in rental fees depending on the source material.
The total fivefold drop in cost came from both newer algorithms and better engineering practices like feeding pages more evenly and using smaller scans. Van Strien noted that optimisation saved as much money as switching to newer models did, if not more. Accuracy remained consistent across the board, though some older systems still hallucinated text on blank pages or repeated content unnecessarily. These figures apply strictly to the specific GPU rental and test parameters described.
- Cost per 1,000 pages fell from 51 cents to under 10 cents
- Optimisation contributed significantly to the total savings
- Test set included 1771 Encyclopaedia Britannica pages




