OCR models now cost under 10 cents per 1,000 pages, down from 51 cents in 18 months
Daniel van Strien of Hugging Face compared eight text-recognition models on 850 historical pages. Cost per 1,000 pages fell from about 51 cents (RolmOCR, April 2025) to just under 10 cents (LightOnOCR-3-0.8B, October 2026), and a million pages now runs between $65 and $130 in GPU rental.
Digitizing a million pages with freely available OCR models now costs between $65 and $130 in GPU rental, depending on the source material, according to a comparison by Daniel van Strien of Hugging Face. The cost of the newest model tested is about a fifth of the oldest. The comparison was reported by The Decoder on November 19, 2026.
Test setup
Van Strien compared eight models that recognize text in images. He ran them on 850 pages drawn from three kinds of material: old natural history books, the 1771 Encyclopaedia Britannica, and typed and printed archival documents.
All models ran on the same GPU, an Nvidia RTX PRO 6000, which van Strien rented through Hugging Face for $2.75 per hour.
Cost results
At the same image size, the oldest model in the test, RolmOCR (April 2025), cost about 51 cents per 1,000 pages. The newest, LightOnOCR-3-0.8B (October 2026), came in at just under 10 cents per 1,000 pages. The five models between them are almost on par with each other on cost.
The $65 to $130 range for a million pages reflects GPU rental cost across the different source materials.
Optimization mattered as much as model choice
According to van Strien, optimizing how the models ran saved about as much money as switching to newer models did, if not more. The optimizations he lists:
- Feeding pages to the model more evenly
- Using smaller scans
- Updating software
- Running a faster GPU
Accuracy and failure modes
The newest model read book pages at least as accurately as the oldest one. Van Strien also found failure modes in some models: they hallucinated text on blank pages or repeated the same content over and over.
What this means
For teams digitizing archives, the measured GPU cost per page is now low enough that rental cost is a small line item at the scale tested: under 10 cents per 1,000 pages for the newest model on this hardware. Those figures cover GPU rental on one rented card at $2.75 per hour, not a full production pipeline.
The finding that runtime optimization saved about as much as a model upgrade, if not more, means that choosing a newer model is only part of the cost equation. Batching, scan size, software versions and GPU choice all moved the result in van Strien's tests.
The accuracy notes are a practical caution. Hallucinated text on blank pages and repeated content are failures a cheaper model can introduce silently, so output checks on blank and sparse pages belong in any large digitization run. The newest model matched the oldest on book-page accuracy, so the cost drop did not come with an accuracy loss on that material.
Related Articles
Hugging Face releases Qwen3.8-27B-pi-adapted LoRA, lifting SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7%
Hugging Face released Qwen3.8-27B-pi-adapted, an Apache-2.0 LoRA adapter trained with reinforcement learning for the pi coding agent. According to MarkTechPost, it raises SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.
InclusionMed launches medical AI evaluation platform, with a 22-model first index across 1,034 tasks
InclusionMed, which describes itself as the world's first independent evaluation and analysis platform for healthcare AI, has gone live with a nine-cell benchmark map (InclusionMed Atlas) and a first composite index (InclusionMed Index). The first round compares 22 models from 11 vendors across five benchmarks and 1,034 tasks.
Databricks launches public AI Index: open-weight models reach 30.9% of enterprise tokens, up from 7.3% in July
Databricks has launched the AI Index, a public resource combining usage data from 10,000+ companies on Unity Gateway with weekly coding-agent evaluations. Open-weight models grew from 7.3% of enterprise token traffic in July to 30.9% in November, and cost per task varies far more than token prices suggest.
NVIDIA Vera Rubin NVL72 posts fastest time to train on six of seven MLPerf Training v6.1 benchmarks
In its first MLPerf Training appearance, NVIDIA's Vera Rubin NVL72 delivered the fastest time to train on six of seven benchmarks, with up to 2x the per-GPU performance of an optimized GB300 NVL72 baseline, according to NVIDIA. On DeepSeek-V3 671B it trained in 4.14 minutes on 1,024 GPUs versus 8.03 minutes for GB300 NVL72.
Comments
Loading...