Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching models
Document extraction is now cheap enough that serving setup, not model choice, decides much of the bill.
What happened
Daniel van Strien of Hugging Face ran eight open OCR models released between April 2025 and October 2026 on the same 850 pages: historical natural history books, the 1771 Encyclopaedia Britannica, and typed and printed archive documents. Every model ran behind a vLLM server on one Nvidia RTX PRO 6000 rented at $2.75 an hour.
Reading 1,000 pages fell from $0.51 with RolmOCR, from April 2025, to $0.095 with LightOnOCR-3-0.8B, from October 2026. With the newest model, he estimates a million pages at roughly $65 to $130 of GPU time, about a day and a half on one GPU for a mixed set.
Quality held on ordinary pages. On book pages every model got more than 99% of characters right on a typical page. On dense 18th-century print, only the newest model roughly matched the April 2025 one. The Decoder notes some models hallucinated text on blank pages or repeated content.
The biggest finding is about operations. For one model on one page type, keeping a vLLM server continuously fed, a larger GPU, 2,048 pixel images and newer vLLM cut cost about four times. Switching to the newest model cut it about three times more. His earlier script left the GPU idle about a quarter of the time.
My take
For businesses with invoices, contracts or intake forms, this changes the build versus buy question. At these prices, running an open model on rented GPUs is realistic for large backlogs, and the data stays under your control.
The serving lesson applies to any batch AI job. If your pipeline sends fixed batches and waits, you're paying for idle hardware. Keeping the queue full is usually a smaller change than a model migration.
I'd also take the blank page finding seriously. In a real pipeline, an OCR model inventing text on an empty page can create a fake record in a CRM or accounting system. Add a check for blank or near-empty pages and flag low-confidence output for review before anything writes to the system of record.
His test covered archival material, not modern business forms, so test on a sample of your own documents first. The full comparison cost him about $6.
More posts
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
- Meta says a confirmation instruction meant for specific actions landed on every Muse message and caused hallucinated bank balancesNov 20, 2026
