OCR a scanned book with ChatGPT, Claude or Gemini
On 7 Nov 2026 we asked each one to clean and OCR the same scanned book. Gemini's app refused. Claude finished the book but dropped a full-width drawing and read fewer words correctly. ChatGPT Plus did the whole book well. All three tests used a small 10 MB copy; real scans are often far larger.
Key takeaways
- Same book, same prompt, 7 Nov 2026: Gemini's app said it had no tools for the job. Claude and ChatGPT both wrote their own clean-up code and returned a PDF.
- Claude removed a full-width drawing on page 35 of the 60-page test. ChatGPT and FixMyScan kept it.
- Words read in order, against the archive's own text: ChatGPT 92–94%, FixMyScan 91–93%, Claude 87–88% on 500 pages.
- We could only attach a 10 MB copy. Our full-resolution copy of the same 500 pages is 70 MB, and 1,000 pages came to 116 MB, past Gemini's 100 MB upload limit.
The test
We wanted a fair answer to a question readers ask us: can’t a chat AI just do this now? So we gave each one the same job on the same day.
- The book: Ernest Thompson Seton’s Two Little Savages, a scan from archive.org with its text layer removed, so every page was a picture only. The book has over three hundred drawings, which makes it a fair test of what gets kept.
- The prompt, word for word in each app: “This is a scanned book. Please make it Kindle-ready yourself, without any outside service or connector: crop the margins, straighten the pages, turn it into clean black and white, and add a searchable OCR text layer. Keep every page, and give me the finished PDF to download.”
- The accounts: ChatGPT Plus, Claude on a paid plan, and the Gemini app. All on 7 Nov 2026.
- Two sizes: the first 60 pages, then all 500.
What each one did
| Gemini app | ChatGPT Plus | Claude | FixMyScan | |
|---|---|---|---|---|
| 60 pages | Refused; suggested other tools and summarised the book | All 60 pages, OCR 94.2% | All 60 pages, OCR 89.8%, drawing on page 35 removed | All 60 pages, OCR 94.4% |
| 500 pages | Not run | All 500 pages, OCR 92–94% in every 100-page block | All 500 pages, OCR 87–88% | All 500 pages, OCR 91–93% |
| Margins | — | Cropped tight | Cropped | Left uncropped in this run |
The OCR figure is the share of the archive’s own words found, in order, in each result. ChatGPT cropped tighter than our run, which had automatic cropping switched off, so its text sits larger on a Kindle screen.
Gemini
Gemini’s app answered that it didn’t have the tools to edit and return a PDF. It suggested ScanTailor and OCRmyPDF, then offered a summary of the book. Gemini can read a PDF you attach. In this test it would not hand back a cleaned one.
Claude
Claude wrote Python with OpenCV and Tesseract in its own sandbox, ran it, looked at its own pages and fixed what it saw. It kept the blank pages and turned sideways pages upright. Two things went wrong. On page 35 a full-width drawing of a leaf was treated as dirt and removed. And across 500 pages it matched 87–88% of the archive’s words, a few points below the other two.
ChatGPT
ChatGPT took the same route as Claude, OpenCV and Tesseract in a sandbox, and did it well. It kept every plate, cropped the margins tighter than we did, and its OCR matched the archive text as well as ours or slightly better. If your file is small and you already pay for ChatGPT, it is a real option.
Where a chat AI stops
The test had a catch we have to be honest about. We could only attach a 10 MB copy of the book, made at low resolution in black and white. Our full-resolution, picture-only copy of the same 500 pages is 70 MB, and the 1,000-page version is 116 MB. Scans made for libraries and archives are usually like that: greyscale or colour, at high resolution.
Here is what each app says it accepts, checked on 7 Nov 2026:
| App | Largest file | Other limits |
|---|---|---|
| Gemini app | 100 MB | Up to 10 files per prompt |
| Claude | 500 MB | PDFs up to 1,000 pages |
| FixMyScan | 1 GB | Books up to 1,000 pages |
A file the app accepts is not the same as a file its sandbox can clean in one go, and we haven’t yet run the full-size scans through ChatGPT or Claude. We will add those results here when we have them.
What FixMyScan does that a prompt doesn’t
- The same steps for every book. One crop for the whole book, the same clean-up and the same OCR, so a shelf comes out alike.
- Pages to check are marked. Blank pages, odd crops, unsure readings and picture pages are flagged, and any page can go back to the original.
- Pictures stay. Pages with photos, maps or engravings are kept in greyscale while the text pages turn black and white.
- DjVu and spreads. It reads DjVu directly and splits two-page scans in reading order.
Which to use
| Your file | Best fit |
|---|---|
| A short, light scan, and you already pay for ChatGPT | ChatGPT can do it; check the pages yourself |
| A long book, over 100 MB, or full of plates | FixMyScan |
| DjVu, or two pages on every scan | FixMyScan |
| A shelf of books you want done the same way | FixMyScan |
| You want to do it from inside your chat | FixMyScan in your AI app |
Questions
Can ChatGPT OCR a PDF?
On a paid plan, yes. In our test ChatGPT Plus wrote its own clean-up and OCR code, ran it on all 500 pages and handed back a searchable PDF. It needs a file small enough to attach, and you check the result yourself.
Can Claude OCR a scanned PDF?
Yes, with code execution turned on. Claude finished both test books, but read fewer words correctly than ChatGPT or FixMyScan and removed one illustration. Claude's upload page sets 500 MB per file and 1,000 pages per PDF.
Can Gemini OCR a PDF?
Gemini can read a PDF and summarise it, but in our test its app would not produce a cleaned, searchable file. It suggested ScanTailor and OCRmyPDF instead. Its upload limit is 100 MB per file.
How did you score OCR?
The archive.org copy of the book has its own text layer. We counted how many of its words appear, in order, in each output. It measures agreement with that layer, not perfect truth, and every tool was scored the same way.
Why not just use ChatGPT then?
For one small, clean file you already pay for, it can work. FixMyScan is built for the books a chat can't take: scans over 100 MB, DjVu files, two-page spreads, plate-heavy books, and a whole shelf done the same way every time.
Can I use FixMyScan inside Claude or ChatGPT?
Yes. FixMyScan connects to AI apps, so you can ask for a book to be fixed from the chat. See FixMyScan in AI apps.