The comparison and the method
A comparative test conducted by Daniel van Strien, a researcher at Hugging Face, measured how much it costs today to extract text from scanned materials using freely accessible artificial intelligence models. The sample comprised 850 pages: old natural-history volumes, the 1771 edition of the Britannica, and typewritten and printed archival documents. The news was reported by IT之家.
All eight models examined were run on the same graphics card, an Nvidia RTX PRO 6000 rented through Hugging Face at a cost of $2.75 per hour. Keeping hardware conditions identical makes it possible to attribute spending differences to the model and to the process configuration, not to the infrastructure.
The collapse in costs
With image size held constant, the RolmOCR model, published in April 2025 and the oldest in the group, required about $0.51 per thousand recognized pages. The newest model, LightOnOCR-3-0.8B, released in October 2026, falls below $0.1 for the same number of pages. At a larger scale, processing one million pages would entail a GPU rental of between $65 and $130, with variations tied to the type of original material.
The decrease is not solely an effect of changing models. According to van Strien, optimizing how models are run can save as much as, or more than, switching to a newer version. The measures indicated include a more balanced distribution of page-processing tasks, the use of smaller scans, software updates, and the use of faster GPUs.
Accuracy and unresolved limitations
On the quality front, the newest model, when applied to book pages, does not appear inferior to the oldest. However, documented problems remain: some models tend to generate text on blank pages or to repeat the same content several times. These are significant errors for those digitizing historical archives, where transcription reliability matters as much as cost.
The main finding concerns processing costs on a limited sample and under controlled hardware conditions. The material does not indicate how the results transfer to other types of documents, to different languages, or to different scan qualities, nor does it specify which of the five intermediate models were compared. The cost reduction is therefore confirmed within the scope of the test, while its generalization remains to be verified.
Why it matters
The digitization of books and archives has long been an expensive activity, often limited by the availability of resources. If recognizing a thousand pages falls below $0.1, the economic constraint shifts from computing power to human verification of errors. The news, attributed to IT之家 on the basis of van Strien's work, signals a measurable change, but not yet a complete solution.
