Hugging Face's van Strien finds open OCR now costs $65 to $130 per million pages
In his test of eight open models, the newest, LightOnOCR-3-0.8B, cost just under 10 cents per 1,000 pages, versus about 51 cents for April 2025's RolmOCR.
What happened
OCR means software that reads text from images of pages, such as scans of old books. Daniel van Strien of Hugging Face compared eight free AI models that do this job. The cheapest run now costs far less than the oldest one did.
Why it matters to you
Libraries, archives and companies with piles of paper records can turn them into searchable text for much less money. Some models still invent text on empty pages, so results need checking.
The takeaway
Digitizing large paper collections with free AI models is now cheap, but output still needs checking.
What it means for you
Runtime tuning such as batching, image size, software versions and GPU choice can cut cost as much as swapping models, and blank-page hallucination and looping need testing.
Bulk digitization of archives and records now costs roughly $65 to $130 per million pages in GPU rental, according to van Strien's tests.
The reported fivefold drop in 18 months in open-model OCR cost per page is a data point on falling inference prices; the source gives no company financials.