What happened
Open AI models that turn scanned pages into machine-readable text have become roughly five times cheaper to run over about 18 months, The Decoder reported. The finding comes from Daniel van Strien of Hugging Face, who benchmarked eight freely available text-recognition models on 850 pages drawn from old natural history volumes, the 1771 Encyclopaedia Britannica, and typed and printed archival records.
Every model ran on identical hardware, an Nvidia RTX PRO 6000 rented through Hugging Face at $2.75 an hour. At a matching image size, RolmOCR, the oldest model tested, from April 2025, cost about 51 cents per 1,000 pages. LightOnOCR-3-0.8B, from October 2026, cost slightly under 10 cents. The five models released between them landed close to one another. Processing a million pages now costs $65 to $130 in GPU rental, depending on the material.
Van Strien said tuning how the models were run cut costs by about as much as adopting newer models, and possibly more. The changes included feeding pages through more evenly, using smaller scans, updating software and moving to a faster GPU. On book pages, the newest model was at least as accurate as the oldest. Some models, however, invented text on blank pages or got stuck repeating the same content.
Context
The benchmark covered only freely available models, released between April 2025 and October 2026, that extract text from images of scanned books and records. All of them ran on a single rented GPU type. The test mixed historical printed books with typed and printed archival documents, and costs varied with the source material. The quoted figures reflect GPU rental time only.
الموجز
أسبوع التقنية في رسالة واحدة.
التقارير التي تستحق وقتك، كل أسبوع. يمكنك إلغاء الاشتراك في أي وقت.
