Models

Digitizing a million scanned pages now costs under $130 as OCR prices collapse

Hugging Face's Daniel van Strien tested eight free OCR models on 850 archival pages: per-1,000-page costs fell from 51 cents to under 10 cents in 18 months, and pipeline tuning saved as much as new models.

OCR models now digitize a million pages for under $130 as costs drop fivefold in 18 months
OCR models now digitize a million pages for under $130 as costs drop fivefold in 18 monthscongerdesign / pixabay
By Elena Vasquez6 min read

Why it matters

  • OCR costs dropped fivefold in 18 months: from about 51 cents per 1,000 pages (RolmOCR, April 2025) to just under 10 cents (LightOnOCR-3-0.8B, October 2026).
  • A million pages now digitizes for $65–$130 in GPU rental, depending on source material.
  • Daniel van Strien of Hugging Face compared eight free OCR models on 850 pages, including the 1771 Encyclopaedia Britannica, on a rented Nvidia RTX PRO 6000 at $2.75/hour.
  • Optimizing the runtime — balanced page feeding, smaller scans, software updates, faster GPU — saved as much money as switching to newer models, if not more.
  • The newest model matched or exceeded the oldest on accuracy, but some models hallucinated text on blank pages or repeated content.

OCR costs have fallen fivefold in 18 months, and a million scanned pages can now be digitized for between $65 and $130 in GPU rental time. That is the headline finding from Daniel van Strien of Hugging Face, who ran a side-by-side comparison of eight freely available AI models that recognize text in images and published the results of what they cost and how they behaved on real archival material.

The numbers matter because they sit at the intersection of two ongoing stories in AI: the collapse in inference costs and the quiet revival of OCR as a practical, open-tooling workload. Institutions holding millions of scanned pages — libraries, archives, government record offices — have historically faced per-page pricing or labor-intensive manual correction. A benchmark that prices a million pages at double digits changes what digitization projects can even contemplate.

What exactly did van Strien test?

The test corpus was deliberately unglamorous: 850 pages drawn from old natural history books, the 1771 Encyclopaedia Britannica, and typed and printed archival documents. This is the messy middle of scanned text — antique typography, aged paper, inconsistent layouts — not clean modern PDFs.

All eight models ran on identical hardware. Van Strien rented a single Nvidia RTX PRO 6000 through Hugging Face at $2.75 per hour, which means the cost comparison reflects model behavior rather than differences in infrastructure. Everything else was held constant so the price-per-page figures would be comparable.

The models span an 18-month window. The oldest is RolmOCR, released in April 2025. The newest is LightOnOCR-3-0.8B, released in October 2026. Six other models sit in between, and van Strien's data shows the five models between the two endpoints cost almost the same as each other — the meaningful movement happens at the newest generation, not in the middle of the timeline.

How steep was the price drop?

At the same image size, RolmOCR processed 1,000 pages for about 51 cents. LightOnOCR-3-0.8B processed the same volume for just under 10 cents. That is the fivefold cost reduction across the 18 months between the two releases, measured on identical rented hardware and identical source material.

Scaled up, the economics are stark. Van Strien puts a full million pages at $65 to $130 in GPU rental, with the range depending on the source material — dense or degraded scans cost more to process than clean ones. For archives sitting on backlogs measured in millions of pages, that price band turns digitization from a capital project into a line item.

What mattered more: newer models or better plumbing?

Here is the finding that practitioners should not skip. According to van Strien, optimizing how the models ran saved about as much money as switching to newer models did — if not more.

That is a significant claim, and it reframes the story. The fivefold drop from April 2025 to October 2026 is not purely a function of better model weights. Van Strien's own engineering choices bought a comparable share of the savings. His optimization levers were concrete and low-tech:

  • Feeding pages to the GPU more evenly, avoiding idle time
  • Using smaller scans where the resolution allowed it
  • Updating the software stack the models ran on
  • Running on a faster GPU

None of those steps require a research lab. They require attention to throughput, and van Strien's results suggest that a team running year-old models with tuned pipelines could match or beat a team running the newest model on a sloppy setup. For budget-constrained archives, that is arguably the most actionable finding in the comparison.

Did cheaper come at the cost of accuracy?

On the core question, no. Van Strien reports that the newest model, LightOnOCR-3-0.8B, read book pages at least as accurately as the oldest one, RolmOCR. The cost collapse did not trade quality for speed — the newest generation holds the accuracy line while cutting cost per 1,000 pages from roughly 51 cents to under 10 cents.

The failure modes are elsewhere. Some models in the comparison hallucinated text on blank pages, producing invented content where the source had nothing to read. Others repeated the same content over and over, looping output instead of transcribing what was actually on the page. These are known behaviors in vision-language models applied to OCR, and van Strien's benchmark makes their prevalence visible across the current free-model field.

For production digitization, that distinction matters. A model that is cheap and accurate on populated pages but fabricates text on blanks can silently corrupt an archive — the errors do not announce themselves. Van Strien's findings suggest that anyone deploying these tools at scale needs blank-page handling and repetition detection in the pipeline, not just a cost-per-page estimate.

Why does a fivefold drop in 18 months matter?

Two forces are compounding here. First, the models themselves are free — van Strien tested freely available AI models, so the only cost is compute. Second, the compute cost is itself falling fast enough that the price of digitizing a page has become almost negligible at the newest generation.

The 18-month timeline is worth pausing on. RolmOCR dates to April 2025; LightOnOCR-3-0.8B to October 2026. In that window, per-1,000-page costs went from about 51 cents to just under 10 cents, and the five intervening models clustered near parity with each other. The pattern — a stable plateau followed by a sharp step down at the newest release — is one that has played out across other AI workload categories, and OCR now has public, reproducible numbers documenting it.

The reproducibility is part of the value. Van Strien ran everything on one rented RTX PRO 6000 at $2.75 per hour through Hugging Face's own infrastructure, on a public corpus of 850 pages from named sources. Other teams can sanity-check their own costs against those figures rather than trusting vendor pricing sheets.

What does this mean for archives and digitization budgets?

The practical read for anyone holding scanned collections is that the compute cost of OCR has stopped being the bottleneck. At $65 to $130 per million pages, even a large digitization program's GPU bill is modest. What remains expensive is everything around it: preparing material, validating output, and catching the hallucination and repetition failures van Strien documented.

The benchmark also tilts the calculus toward doing optimization work in-house. Since van Strien found that pipeline tuning saved as much as or more than upgrading models, institutions get more mileage from an engineer who balances GPU load and right-sizes scans than from chasing every new model release. The newest model helps; it is not the whole story.

The direction of travel is clear from the data. If costs fell fivefold between April 2025 and October 2026 while accuracy held or improved, the next generation of free OCR models will push per-page costs lower still — and the binding constraints on mass digitization will shift almost entirely to verification, curation, and the handling of the failure modes this comparison exposed.

Original: danielvanstrien.xyz

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

744 articles

Related articles

  1. OpenAI's textGrain Watermark Falters on Short and Edited Text
  2. OpenAI Releases GABRIEL, an Open-Source Toolkit for Measuring Qualitative Data with GPT
  3. OpenAI launches GPT-6.2 Astra and Sol with 20 percent lower API prices
  4. Oxford lets OpenAI train AI models on Bodleian Library texts
  5. OpenAI's SWE-Lancer Asks If LLMs Can Earn $1 Million Freelancing

« Previous article