10 Cents per 1,000 Pages: How Well Can Cheap Open-Source OCR Really Read Fraktur?
Open OCR models can now read old books for less than ten cents per 1,000 pages. But the well-known tests leave out the very typeface used for German books, newspapers and official gazettes until 1941. We gathered the Fraktur measurements that actually exist and looked at what they mean for archives and genealogists.

What's going on
On November 18, 2026, Daniel van Strien of Hugging Face calculated just how cheap text recognition has become. OCR (optical character recognition) is the technology that turns a scanned image of a book page into searchable text. According to van Strien, the newest open model he tested, LightOnOCR-3-0.8B from October 2026, costs around $0.095 per 1,000 pages. RolmOCR from April 2025 cost about $0.51. The figures are based on renting an Nvidia RTX PRO 6000 on Hugging Face Jobs at $2.75 per hour. At that rate, a million pages would cost $65 to $130 in compute. The news was picked up internationally, including by the Chinese portal IT之家.
For German libraries, archives and genealogists, one question is still open. Until it was banned in 1941, much of Germany's printed output, from books and newspapers to official gazettes, was set in Fraktur. This blackletter typeface has the long ſ and many ligatures (joined letter pairs like “ch” or “ck”). It's exactly the typeface missing from the widely cited tests. So we gathered the Fraktur measurements that do exist and asked: do the budget models read Fraktur just as well, or do they quietly rewrite the text?
The blind spots of the well-known tests
Van Strien's cost test covers 850 pages: natural history books, the 1771 Encyclopaedia Britannica and typed archival documents, mostly from the late 20th century. He doesn't mention Fraktur or German-language pages. He lists plenty of limitations himself: the sources are few, and the “five times cheaper” figure rests on a single model pair at a single image size. Even on the Britannica, several models read the long ſ as f. On blank pages, GLM-OCR described the image in a sentence every time, and OvisOCR2 answered “The quick brown fox” 24 times.
FineBooks, the leaderboard from Hugging Face and EleutherAI, tests 14 open models on 2,165 pages in English, French, German and Latin. So German is included, but only in Antiqua, the roman typeface still used in print today. The blog post says explicitly that the evaluation tells you nothing about German Fraktur, handwriting or non-Latin scripts. There's no breakdown by language. The leader is dots.mocr, with 97.6 percent character accuracy at $1.94 per 1,000 pages. PaddleOCR-VL-1.6 reaches 96.1 percent at $0.34. The authors themselves warn that this isn't yet good enough for scholarly transcription, because the models silently modernize the long s and ligatures.
What Fraktur data actually exists
| Dataset | Material | Size | Long ſ | Open OCR specialist models tested? |
|---|---|---|---|---|
| RISE Fraktur Adverts (University of Basel) | Basler Avisblatt, 1731 and 1762 | 5 pages, 93 scored ads | historical spelling preserved | no, only general-purpose models |
| HIPE-OCRepair dta19 | 39 books from the Deutsches Textarchiv, 18th/19th c. | 30 test pages per noise level | normalized in the reference text | no, tests post-correction |
| HIPE-OCRepair impresso-nzz | NZZ front pages 1780–1947 | 150 training, 17 test pages | – | no, not scored |
| Reichsanzeiger-GT (UB Mannheim) | Reichsanzeiger 1820–1939 | 197 pages, 119,429 lines, CC0 | transcribed as ſ | no published results found |
The RISE benchmark from the University of Basel is the only actively maintained Fraktur comparison. Models there don't just have to read ads; they also have to number them and assign them to sections. The score is based on the character error rate (CER, the share of misread characters). If a model misses an ad, the whole ad counts as wrong. So a low score can come from getting the structure wrong, not just from misreading. At the top are paid models like Claude Opus 5.5 and Gemini 4 Argon, at 99.4 to 99.5 points. Among open-weight models, the field looks like this:
| Open model (per RISE list) | Score |
|---|---|
| Kimi K3 | 98.4 |
| Qwen3.5-122B-A10B | 97.0 |
| Qwen3.5-27B | 95.1 (another run: 78.2) |
| GLM-5.3-Flash | 91.8 |
| Gemma 4 31B | 60.7 |
| Qwen3.5-9B | 49.7 |
| Ministral 3B | 44.6 |
The actual budget models from van Strien's calculation (LightOnOCR, PaddleOCR-VL, GLM-OCR and dots.ocr) don't appear on the list at all. Also notable: once models drop below roughly 30 billion parameters, performance collapses.
HIPE-OCRepair (ICDAR 2026) measures something else: how well language models fix faulty OCR after the fact. The DTA pages were recognized with Tesseract's Fraktur model. According to the authors, the result was so good (0.4 percent errors) that they had to inject artificial noise of 3 and 7 percent. The best team (BnF, with a fine-tuned 24-billion-parameter Mistral Small 3) cut the error rate from 5.5 to 0.8 percent. On near-flawless input, correction barely helped, and the authors warn about overcorrection. Here's the catch for our question: the DTA reference text normalizes the long ſ. So this benchmark can't show whether a model modernizes it.
Real-world tests: big models win, small ones are unproven
The most concrete comparison we found comes from an open GitHub project built for a Harvard lecturer, dated November 5, 2026. It tested real pages from Grimms' Fairy Tales (1812) and Die Gartenlaube (1853):
| Reader | Grimm 1812 | Gartenlaube 1853 | Cost per 1,000 pages (converted) |
|---|---|---|---|
| Claude Haiku 5.5 | 0.9% | 1.5% | approx. $0.90 |
| Gemini 3.8 Flash | 1.0% | 1.4% | approx. $9.90 |
| Qwen3-VL-235B (open) | 1.1% | 1.7% | approx. $1.90 |
Kraken german_print |
2.9% | 3.4% | local, free |
Tesseract frak2021 |
9.3% | 7.4% | local, free |
The values are “normalized,” meaning ſ and s count as the same letter. On the Antiqua page (Schiller, 1796), Kraken came out on top. According to the README, Claude Haiku turned ſ into s in every line, even though the prompt told it not to. With only eight pages, that's a hint, not proof.
The notes from the Source Library project (Embassy of the Free Mind), covering October 31 to November 18, 2026, are more sobering. There, olmOCR had 2.6 percent errors on 21 Fraktur pages, while Gemini Flash-Lite managed 0.5 percent. PaddleOCR-VL did worse than Flash-Lite, and LightOnOCR's output was accepted for only 2 of 18 pages. The team therefore decided against fine-tuning vision-language models (VLMs, AI models that process images and text together) for Fraktur. Their reason: there's “no evidence” that a VLM can read Fraktur. A 2025 discussion about the Danish kb-books project made a similar point: LLM-OCR errors look plausible, which makes them hard to spot.
Reading or rewriting?
The Amazon paper “Do VLMs Read or Rewrite?” (2026) put a number on the risk. The researchers deliberately altered words in test documents. Then they counted how often a model changed them back to the “correct” form instead of reading what was actually on the page:
| Model | Share of rewritten words |
|---|---|
| Tesseract, Gemini 3 Flash | 0% |
| PaddleOCR-VL-1.5 | 7.9% |
| DeepSeek-OCR-2 | 13.5% |
| Qwen3.5-4B | 27.4% |
| LightOnOCR-2-1B | 53.7% |
| olmOCR-2-7B | 58.6% |
The tests covered English, Chinese and Korean on synthetically generated pages, not Fraktur. For historical digitization, the authors recommend classic OCR or carefully chosen specialist models. Our take: Fraktur is the worst-case scenario for these models. To a model with a strong feel for language, historical spellings like “Thür,” “seyn” or “Urtheil” look like typos. LightOnOCR is the cheapest model in van Strien's calculation, and its previous version was among the worst rewriters. That's a red flag. Version 3, however, hasn't been tested for this yet.
We didn't run our own test with van Strien's ocr-bench tool for this article. If you want to try it yourself, keep in mind that ocr-bench doesn't measure an error rate against a reference text. Instead, an AI judge (Qwen3.8-27B by default) compares two outputs at a time against the page image. In the Danish discussion, participants noted that such judges often mark pre-1800 spellings as errors. For a fair Fraktur test, UB Mannheim's Reichsanzeiger-GT is a better choice, because it keeps the long ſ and is released under CC0.
What the classic tools cost
Transkribus offers models trained specifically for Fraktur. The University of Zurich's “German Fraktur 18th–20th century” model, trained on NZZ front pages from 1780 to 1940, has a 0.5 percent error rate on its own test pages. UB Mannheim's “GermanNewspapers-M1” reaches 0.7 percent. Each figure applies only to that model's own test data, so they aren't directly comparable. According to the FAQ, printed text costs 0.5 credits per page, though the pricing page only lists the rate for handwriting. By our calculation, that works out to about €55 per 1,000 pages with the Scholar package (€99 for 900 credits). With top-up credits (€59.50 for 250), it's about €120. That's more than 500 times van Strien's $0.095. Kraken, Calamari and OCR4all, by contrast, run for free on your own machine. Published studies report error rates as low as 0.18 percent for well-trained Calamari models, but only on matching material (DTA19).
Our verdict
The 10-cent figure is accurate, just not for Fraktur. No published Fraktur measurement makes a case for the small, cheap OCR specialist models. The few real-world reports put them well behind large models and trained classic tools. Only the large open general-purpose models like Qwen3-VL-235B, Qwen3.5-122B or Kimi K3 read Fraktur well, and they won't run on an ordinary graphics card.
Our recommendations by use case:
- Books and newspapers from 1750–1940, with an ordinary graphics card or CPU only: Kraken with a Fraktur model, or OCR4all/Calamari. These tools don't rewrite anything, can output line coordinates and run without a GPU.
- With a 24 GB card: Qwen3.5-27B scored 95.1 points in one RISE run but only 78.2 in another. Be sure to check it against your own pages.
- Maximum accuracy for newspapers: Transkribus with a newspaper Fraktur model, or a large model via API, with spot checks.
- Church registers: These are usually handwritten in Kurrent, so OCR models for printed text won't help. Your options are Transkribus or the community fine-tune LightOnOCR-3-1B-Historic-German-Handwriting. It has a 4.7 percent error rate and takes about 7 seconds per page on an RTX A4000. According to its author, it's “unusable” on CPU alone.
The most important test is still missing: a clean comparison of LightOnOCR-3, PaddleOCR-VL-1.6, GLM-OCR and dots.mocr on Fraktur reference texts that keep the ſ. The data for it is free to download.
Sources
- Daniel van Strien: OCR cost trend (18.11.2026) danielvanstrien.xyz
- IT之家: Kosten der Buchdigitalisierung ithome.com
- Hugging Face Blog: FineBooks Historical Books OCR Leaderboard huggingface.co
- The Decoder: FineBooks the-decoder.com
- RISE Humanities Data Benchmark: Fraktur Adverts rise-services.rise.unibas.ch
- arXiv 2607.08143: ICDAR 2026 HIPE-OCRepair arxiv.org
- arXiv 2607.21617: Do VLMs Read or Rewrite? arxiv.org
- GitHub: UB-Mannheim/reichsanzeiger-gt github.com
- GitHub: davanstrien/ocr-bench github.com
- GitHub: mueser-prototypes, Fraktur-Bake-off github.com
- GitHub-Issue #7845: Fraktur-Tests offener Modelle (Source Library) github.com
- Hugging Face Diskussion: chcaa/kb-books OCR-Modelle huggingface.co
- Transkribus: Preise transkribus.org
- Transkribus: Subscription FAQ transkribus.org
- Transkribus: Modell German Fraktur 18th-20th century transkribus.org
- Hugging Face: LightOnOCR-3-1B-Historic-German-Handwriting huggingface.co
- Hugging Face Blog: Can LLMs Correct Historical OCR Without Rewriting History? huggingface.co
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.