Across an evaluation benchmark of 850 pages from old natural history books, the 1771 Encyclopaedia Britannica, and typed archival documents, open-source AI models demonstrated a steep drop in text extraction expenses. As Daniel van Strien of Hugging Face stated, comparing eight models that recognize text in images allows businesses to measure processing costs directly against real-world accuracy.
The findings show that pulling text out of scanned books and records has turned from a heavy infrastructure commitment into a marginal operational expense. A million pages now runs between $65 and $130 in GPU rental fees, depending on the source material.
The Drop in Unit Economics
The trajectory of processing costs shows how rapidly open-source model design is compressing operational expenses. On 2026-07-06, Baidu Unlimited OCR tackled scaling document digitization by limiting its attention window to the last 128 units while anchoring visual data, using a 3-billion parameter Mixture-of-Experts scheme with only 500 million active parameters. A separate benchmark on handwritten Russian text on 2026-08-01 showed that large models like the Qwen2.5-VL (7B) fall short in operations compared to specialized architectures like PaddleOCR-VL, which has 1.7B parameters and runs five times faster.
The cost of scanning 1,000 pages has dropped from 51 cents for the oldest model to just under 10 cents for the newest one.
Testing under identical conditions on an Nvidia RTX PRO 6000 GPU rented through Hugging Face for $2.75 per hour reveals this downward pricing slope. At the same image size, the oldest tested model, RolmOCR from April 2025, cost about 51 cents per 1,000 pages. By October 2026, the LightOnOCR-3-0.8B model dropped that expense to just under 10 cents per 1,000 pages, with the five models released between them performing almost on par with each other.
Pipeline Tuning and Quality Trade-offs
Model updates alone do not account for the financial gains. According to Daniel van Strien, optimizing how the models ran saved about as much money as switching to newer models did, if not more.
Optimizing how the models ran saved about as much money as switching to newer models did, if not more.
In practical execution, that optimization meant feeding pages more evenly, using smaller scans, updating software, and running a faster GPU. The latest architecture proved resilient on difficult historical scans: LightOnOCR-3-0.8B read book pages at least as accurately as RolmOCR. However, the technology still presents operational friction. Some models hallucinated text on blank pages or repeated the same content over and over. Deploying open-source text extraction at scale delivers negligible hardware compute bills, but leaves organizations with the task of automated quality filtering to catch phantom outputs.
What doesn’t add up
Running an Nvidia RTX PRO 6000 at $2.75 per hour turns raw text ingestion into a minor expense, but the tendency of open weights to loop or invent text on blank sheets creates an unexpected bottleneck downstream. A similar friction surfaced when a specialized model outperformed the larger Qwen2.5-VL on handwritten text, exposing the limits of relying purely on smaller inference bills. Trimming compute fees down to pocket change merely shifts the invoice toward engineers writing filters for phantom text.
Our stories: OCR Specialization: PaddleOCR-VL vs. Universal Models (August 1).