OCR with open models got about 5× cheaper in 18 months
ocr
glam
huggingface
benchmarks
I ran eight open OCR models from the last 18 months on the same 850 pages and the same GPU. The newest costs about a fifth of the oldest, and the model is only part of the reason.
tl;dr: the newest open OCR model I tested reads 1,000 pages for under $0.10, about a fifth of what an April 2025 model cost, and it reads ordinary pages as well. Running models better saved me more than switching to the newer model did.
The short version
In February I wrote that you could re-OCR a digitised collection for about $0.002 a page. That number is already out of date, so I measured it again (with some help from my agent). I ran eight open OCR models, released between April 2025 and October 2026, on the same 850 pages, on the same GPU, served the same way.
Each dot is one model at the date it came out. Hover a dot for the exact cost, and use the menu to see one kind of page at a time.
- Cost: reading 1,000 pages fell from $0.51 with the April 2025 model to $0.095 with the newest, LightOnOCR-3-0.8B. That is about five times cheaper in 18 months at the same image size.1 Five models from October 2025 to July 2026 cost $0.23 to $0.28 and are too close to rank.2 The fitted trend halves the cost about every ten months, but I would not bet on it.3
- Quality: on ordinary pages the cheaper models did not get worse. On book pages every model gets more than 99% of characters right on a typical page (not counting running headers), and a vision-model judge preferred the newest model to the April 2025 one on book and typed pages. On dense 18th-century print the April 2025 model still beat every later model except the newest, which about ties with it.
- Where the saving came from: for one model on one kind of page, running it better cut the cost about four times: a vLLM server kept busy with many pages at once, instead of a script sending fixed batches, plus a bigger GPU and newer serving software. Swapping in a smaller, faster model cut it about three times more.
- What to use: for book pages and typed documents, start with LightOnOCR-3-0.8B. It is the cheapest on all three kinds of page, and there it read as well as or better than the April 2025 model. For dense historical print, no model is clearly best: try it and the April 2025 model (RolmOCR) on a sample of your own pages first.
- What a collection costs: with the newest model, a million pages is roughly $65 to $130 of GPU time, depending on the kind of page, and about a day and a half on one rented GPU for a mixed set.
The sections below show the evidence. The details are in the appendix.
Did the cheaper models get worse?
On ordinary pages, hardly. Median error rates on the 282 book pages with checked transcriptions run from 0.44% to 0.82%, scored with the text rules from bhl-ocr-eval, the code behind the BHL OCR leaderboard, and the newest model makes no more errors than the April 2025 one. Switch the measure to see the verdicts of a vision-model judge, using the prompt from ocr-bench; on dense print, only the newest model keeps up.4
Where the saving comes from
I took one model, GLM-OCR from January 2026, and one kind of page, the 1771 Britannica. Run the way I ran it in February, it cost $1.57 per 1,000 pages. Then I changed how it runs: a vLLM server that saturate keeps fed with many pages at once, a larger GPU (an RTX PRO 6000 instead of an L4), page images fitted to 2,048 pixels, and a newer vLLM. That brought the same model to $0.40. Swapping in the newest model, LightOnOCR-3-0.8B, which is smaller and faster, brought it to $0.13.5
I got less stupid. In February my script sent pages in fixed batches of 64 and did its own processing between them, so the GPU sat idle about a quarter of the time. Now the server always has pages waiting.
This is the last step in practice: the newest model reading typed pages, several a second.
The on-screen rate in the clip is a live reading.6
What would your collection cost?
Type in the number of pages in your collection and pick the kind of page closest to yours. Every model shows as a tick, so you can see how much the choice of model moves the bill.
Split the pages over several GPUs and the time falls while the cost stays about the same. Treat the result as a rough estimate.7
My advice from February stands, and it is now cheap to follow: run a few models on a sample of your own pages and look. The first full comparison of all eight models on all 850 pages cost about $6.
Each model ran with the settings from its uv-scripts OCR recipe, driven by saturate.
Appendix
NoteHow I ran it
- Models: RolmOCR (April 2025), dots.ocr (July 2025), olmOCR-2 (FP8, October 2025), PaddleOCR-VL (October 2025), DeepSeek-OCR (October 2025), GLM-OCR (January 2026), OvisOCR2 (July 2026) and LightOnOCR-3-0.8B (October 2026).
- Pages: 350 pages from historical natural history books, 250 pages of the 1771 Encyclopaedia Britannica, and 250 pages from 44 typed and printed documents from an archive, mostly from the late 20th century and some from the 2000s. 282 of the book pages have human-checked transcriptions, so I can measure accuracy on those. 63 of the book pages are blank.
- Hardware: one Nvidia RTX PRO 6000 on Hugging Face Jobs, at $2.75 an hour.
- Setup: every model ran behind a vLLM server, with the same client (saturate) sending it many pages at once. Prompts and sampling follow each model’s own documentation where I could manage it. Every page was first fitted to 2,048 pixels on its longest side. Some models then shrink it further by their own rule.
Three models ran outside their usual pipeline. Two are designed to run after a layout-detection step, which I left out. One ran with its plain-text prompt in place of its default.
The shape of the result is two steps with a plateau between them. The two earliest models, from April and July 2025, cost $0.45 to $0.51 per 1,000 pages. Five models released between October 2025 and July 2026 sit between $0.23 and $0.28, too close together to rank. Then one model from October 2026 comes in at $0.095.
NoteThe judge and the blind checks
Two of the cheaper models make about one and a half times as many errors on book pages as the April 2025 model. I measured the eight models twice and the error rates did not move.
I also asked a vision model to compare outputs side by side on 50 pages of each kind. To check that judge, I read eight pairs blind myself, and had four other AI models rate 40 pairs, each checking several differences per page against the scan. Where two of them both picked a side, it was nearly always the same side. The judge itself was less steady. It changed its verdict with the order of display on between one page in ten and one in five, depending on the kind of page, and it agreed least with the other raters on typed pages.
With that caveat, the newest model was clearly preferred to the April 2025 one on book pages: 35 pages against 7. On typed pages it was preferred less clearly: 18 pages against 9, with 22 too close to call.
On the dense two-column print of the 1771 Britannica, the April 2025 model was preferred over every later model except the newest. Several of them read the long s as an f, but that is not the whole story: the blind raters still preferred the April 2025 model on half or more of 20 pairs when they ignored long-s mistakes. The newest model about ties with it: 18 pages against 19.
NoteCost by kind of page
The kind of page moves the cost as much as the choice between neighbouring models, and the dearest kind of page is not the same for every model.
NoteHow models fail
None of this lines up neatly with how new a model is.
- 63 of the 350 book pages in my set are blank. Three models return nothing for a blank page, which is right. GLM-OCR, which I ran without its layout step, writes a sentence about the image instead, on all 63, usually that it is too blurry to read. OvisOCR2 answered 24 blank pages with “The quick brown fox”. RolmOCR writes made-up tables on most blank pages, and on almost half of them repeats an empty table until it hits its output limit.
- On a tax form full of dot leaders, four of the eight models looped in each run, and not the same four both times. The newest model was one of them. No model finished the page cleanly in both runs. Most readings that did finish had mistakes, such as ticking a box that is empty on the form or leaving out lines.
- On a typed page with a black blotch over the last few lines, three models each made up a different final entry, in both runs. Nothing in their output shows that they guessed.
The share of pages where a model hit its output limit ranged from under 1% to 26% by model and kind of page, and the highest rates belong to models I ran outside their usual pipeline (for DeepSeek-OCR, its default prompt cut the book-page rate from 13% to under 1%, at some cost in accuracy).
NoteWhat the first step bundles
My February recipe sent fixed batches of 64 pages and left the GPU waiting at least a quarter of the time. Now a server always has pages queued. The old recipe also sent full-size scans where I now fit pages to 2,048 pixels, and vLLM itself moved on. From partial tests, I would put the way pages are sent, the image size and the engine together at about 2.5 times, and the bigger GPU at about 1.5 times. I could not separate them cleanly.
I also ran three models on the vLLM release that was current when each came out. Today’s version was 1.2 to 1.5 times faster for the two 2025 models, and between level and 1.2 times faster for the 2026 one.
NoteImage size
Sending pages at full scan size instead of 2,048 pixels cost between 1.3 and 4.2 times more, depending on the model and the kind of page, and did not improve accuracy on the book pages. Going the other way, 1,288-pixel images nearly halved the cost of the April 2025 model. But the judge then preferred the larger-image reading on most dense-print pages (36 against 7) and on many book pages (24 against 6). This is why the headline comes with a condition: against the April 2025 model at 1,288 pixels, the newest model is about three times cheaper, not five.
The left panel shows what each image size costs, and the right panel what it does to the error rate on book pages.
NoteSmall models on small GPUs
The newest model ran on a 16 GB T4, at $0.40 an hour, for roughly $0.25 to $0.41 per 1,000 pages. On a 24 GB GPU the April 2025 model loads, but its weights leave room for only a handful of pages at a time, and it cost about $2.10 to $2.80 per 1,000 pages (two kinds of page; the third timed out). A sister model of the same size, with 8-bit weights and smaller images, ran comfortably on the same card. The lowest cost per page is still on the big GPU. The point is that you no longer need one. These small-GPU figures come from short single runs.
NoteWhat this does not tell you
- These are 850 pages from three sources. Your collection is different.
- “About five times” compares one pair of models at one image size.
- Each model was measured twice, in two slightly different ways. For the all-pages cost, differences under about a quarter are not meaningful. For a single kind of page, the bar is nearer a half. The side tests ran once.
- The 2,048-pixel rule is the newest model’s own rule, and my test harness grew out of the recipe for that model. Other models might do better with settings tuned for them.
- Three models ran outside their usual pipeline, and output limits differ by model, from 4,096 to 16,384 tokens.
- The accuracy figures come from 282 book pages. For the other pages I rely on a vision-model judge, which has known biases.
- Newer models here are also smaller, so I cannot separate “newer” from “smaller”.
- I only tested models I could get running today, with recent vLLM versions (0.22 to 0.30).
- Prices are today’s Hugging Face Jobs prices and leave out the few minutes each job takes to start.
- For release dates I used the date each model’s repository was created on the Hub.
NoteAll the numbers
Every model with every main measure. Click a column heading to sort.
Footnotes
If the April 2025 model gets smaller page images (1,288 pixels instead of 2,048), it costs about half as much, and the gap to the newest model is about three times instead of five. But it then reads less well. See “Image size” in the appendix.↩︎
Each model was measured twice, in two slightly different ways. For the all-pages cost, differences under about a quarter are not meaningful. For a single kind of page, the bar is nearer a half.↩︎
The line through the eight points halves about every 9.6 months. Carried on, a million pages would cost under $50 by the end of 2027. It rests on eight models, one of which does most of the work, and with the April 2025 model at 1,288 pixels as the starting point the halving time stretches to about 13 months. Treat it as a toy.↩︎
The judge is a vision model that compared outputs side by side on 50 pages of each kind. It is less steady than I would like. See “The judge and the blind checks” in the appendix.↩︎
The first step bundles several changes: how pages are sent, the image size, the vLLM version and the GPU. See “What the first step bundles” in the appendix.↩︎
The clip peaks at 9.5 pages a second. That run averaged 7.7 pages a second, against 7.6 for typed pages in the benchmark.↩︎
The calculator uses the measured rates on one RTX PRO 6000 at $2.75 an hour. Start-up time is not included. Measurement noise alone is about a quarter either way, and pages unlike mine can differ by more.↩︎