HOT RADAR
toolsHEAT 22.2°

AI CURATED HOT STORY · 11/19/2026

Open Source OCR Costs Drop Fivefold in 18 Months to $130 per Million Pages

2 reports archived2 independent sourcesupdated 11/19/2026, 21:49:04
Synthesis & Latest Updates
2 Sources Cross-Validated

Open source OCR models have cut the cost of digitizing historical texts by about fivefold in 18 months. Hugging Face researcher Daniel van Strien benchmarked eight open source models on the same GPU, finding that processing 1,000 pages fell from roughly $0.51 with RolmOCR in April 2025 to under $0.10 with LightOnOCR-3-0.8B in October 2026, bringing the GPU rental cost for one million pages to just $65–$130.

The test set included 850 real archival pages such as the 1771 Encyclopaedia Britannica. The latest model matches earlier ones in accuracy on book pages but suffers from blank-page hallucinations and repetitive output. The study also notes that engineering optimizations—task scheduling, smaller scan sizes, software upgrades—can deliver savings comparable to or greater than switching to a newer model, and the similar costs of five intermediate models suggest the drop stems mainly from the latest generation rather than gradual improvement.

This trend significantly lowers the barrier to large-scale textification of ancient books and archives, enabling cultural institutions, small libraries, and individual researchers to pursue digitization at minimal cost, potentially accelerating open access and scholarly research on historical documents.

LATEST/The latest benchmark shows LightOnOCR-3-0.8B achieving under $0.10 per 1,000 pages on 850 historical archives, a fivefold drop in 18 months. The study also flags blank-page hallucinations and repetitive output as remaining issues, while engineering optimizations can further reduce costs.

Hourly heat curve

7 fully observed hours
Now
22.2
Peak
22.822:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 2 reports · Latest first
  1. IT之家 · 智能时代T2·72 pts
    • Cost per 1,000 pages dropped from about $0.51 to under $0.1, with a million pages costing roughly $65-130 in GPU time
    • The newest model matches or beats earlier ones on book-page accuracy but still hallucinates text on blank pages and repeats output
    • Pipeline optimizations such as balanced scheduling, smaller images, software updates and faster GPUs can rival the savings of a newer model
  2. The DecoderT2·72 pts
    • Eight open OCR models compared on the same RTX PRO 6000 across 850 real archival pages, including the 1771 Encyclopaedia Britannica
    • Per-1,000-page cost fell from 51 cents to under 10 cents, putting a million pages at $65-$130 in GPU rental
    • The five intermediate models cluster closely, suggesting the drop comes from the newest generation rather than gradual tuning