Daniel van Strien from Hugging Face evaluated eight AI models designed for optical character recognition. The testing process involved processing 850 pages from historical archives, including natural history books and the 1771 Encyclopaedia Britannica. All models were executed on an Nvidia RTX PRO 6000 GPU, which was rented at a rate of $2.75 per hour.
The research highlights a significant reduction in operational costs over an 18-month period. The oldest model, RolmOCR, released in April 2025, processed 1,000 pages for approximately 51 cents. In contrast, the latest model, LightOnOCR-3-0.8B, introduced in October 2026, reduced the cost to just under 10 cents per 1,000 pages.
Current estimates indicate that digitizing one million pages now incurs GPU rental costs ranging between $65 and $130. These figures vary based on the specific source material being processed. The study notes that the newest models maintain accuracy levels comparable to older versions while offering substantial financial savings.
Beyond model selection, technical optimizations contributed significantly to cost efficiency. Improvements included more uniform data feeding, the use of smaller scan sizes, software updates, and the deployment of faster hardware. Despite these gains, some models exhibited performance issues, such as text hallucinations on blank pages or repetitive output.


