OCR is the floor of image translation. A translation cannot be better than the text that was read off the page — misread one character and everything downstream is faithfully translating the wrong input. So it is worth knowing what published benchmarks have actually found.
To be clear up front: none of these numbers are ours. We do not build an OCR engine — recognition and translation run on an upstream model service. What follows is a reading of public benchmarks, each one named and linked so you can check it.
Printed text: the commercial APIs have converged
On clean printed text the major commercial APIs all sit in a high band, and the spread between them is smaller than the spread caused by image quality. A November 2025 roundup put Azure Document Intelligence in front, with GPT-5, Gemini 2.5 Pro, Google Vision and Amazon Textract close behind, all at very high accuracy.
Practically: if your material is clean scans or printed packaging, the engine will not be your bottleneck, and price and integration effort are the more useful things to compare.
Handwriting: this is where they separate
Open-source Tesseract falls off a cliff on handwriting. A study of multi-page handwritten document transcription found it essentially unable to produce a meaningful transcription, while multi-modal LLMs handled the same material far better. That is an architectural gap, not something tuning closes.
Low-resource languages: do not assume coverage
An engine listing a language and an engine being usable in that language are different claims. A comparative study of zero-shot OCR on Sinhala and Tamil found recognition quality well below what the same engines manage on major languages. If your material is in a smaller language, treat the vendor's language list as a starting point and test it on your own files.
Cost and deployment: the price of open source is not the licence
Tesseract is free and runs locally — low latency, and nothing leaves your network. The cost is that deployment, tuning and accuracy recovery are yours. Cloud APIs bill per page and add a round trip, but remove the engineering. Below a certain volume the cloud API is almost always cheaper once your time is counted; self-hosting starts to make sense only well past it.
For image translation, the accuracy number is not the thing that matters most
A general OCR benchmark measures whether characters are read correctly. Image translation also needs reading order (Japanese manga runs vertically, right to left; a webtoon is one enormous vertical strip), the boundaries of bubbles and text blocks, and which fragments belong to the same sentence. None of that is scored in an OCR benchmark, and all of it decides whether the translated result reads properly.
Which is why picking an image translation tool off an OCR leaderboard is optimising the wrong metric. Take the three hardest images you actually have, run them through each candidate, and look at whether the output is usable as-is.
How to choose
- Clean printed documents: any major commercial API will do — decide on price and integration effort.
- Handwriting: avoid Tesseract; use a multi-modal LLM or a dedicated handwriting service.
- Low-resource languages: validate on your own material before committing, whatever the language list says.
- Data cannot leave your network: Tesseract or a self-hosted model, accepting the accuracy cost.
- Image translation: do not shop on OCR scores — judge the end-to-end output.
Sources
