OCRRouter

GPT vs dots.ocr

Pick GPT when you need handwritten notes, custom JSON extraction or document question answering, and you accept a hosted proprietary API. Pick dots.ocr when you want an MIT-licensed model you run yourself that returns layout boxes, tables as HTML and formulas as LaTeX in one pass. The main split is hosted and token-priced versus self-hosted and compute-priced.

GPTopenai/gptdots.ocrrednote/dots-ocr
TypeVision modelOpen source
Price11 credits / page2 credits / page
Per 1,000 pages$11.00$2.00
Typical speed2.4 s / page1.5 s / page
Best forMessy scans, custom instructionsMultilingual document parsing
InputsPDF, imagesPDF, images
Languages50+ languages100+ languages
Max pages per request100500
Tables✓ Yes✓ Yes
Handwriting✓ YesNo
Math✓ Yes✓ Yes
Bounding boxesNo✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ Yes✓ Yes
On OCRRouterAvailableAvailable

Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.

Which to choose

Choose GPT if

  • It handles messy handwriting well: an HN user reported about 5% errors on messy cursive versus about 30% for Mistral OCR.
  • You can prompt it to extract data into custom JSON and ask questions about the document, not only transcribe it.
  • The OpenAI API takes files up to 50 MB per request and sends both extracted text and page images to the model for PDFs.

Choose dots.ocr if

  • It is free to use under MIT-licensed code and you only pay for compute, so bulk OCR cost is predictable.
  • One pass returns layout boxes, categories and text, with tables as HTML and formulas as LaTeX, plus bounding boxes that GPT's standard response lacks.
  • It is a 3B model (1.7B LLM backbone) that you can run on a single CUDA GPU with vLLM, and it targets multilingual PDFs to Markdown.

Key differences

GPTdots.ocr
How you run itHosted OpenAI APISelf-hosted via vLLM or Hugging Face Transformers
LicenceProprietaryMIT (the repository also includes a separate dots.ocr license agreement file)
PriceToken-based per model; page images add input tokens, so cost per page is not fixedFree software; you pay for compute
Output formatsMarkdown, JSON, plain textJSON, Markdown, HTML, LaTeX, bounding boxes, SVG
Coordinates and confidenceNo bounding boxes or confidence scores in the standard responseLayout boxes and categories returned with the text
BenchmarkGPT-5.2: 85.50 on OmniDocBench v1.5 (independent); 86.59 on v1.6 (independent)83.9 on olmOCR-bench for dots.mocr (v1.5), 79.1 for the original dots.ocr (both by the authors); not published on the same benchmark as GPT
Known failureNot publishedRuns of ellipses or underscores can make output repeat endlessly; complex tables and formulas remain difficult; a table above 1000 grid cells raised an error and dropped the page in a Docling issue

Sources

What each one is

GPT

General vision model. Good at messy scans and following extraction instructions.

GPT details

dots.ocr

Single vision model from rednote that does layout detection and content recognition together.

dots.ocr details

Try both on your own file

Run the file once with each model and compare the output.

PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?

Use one, fall back to the other

{"model": "openai/gpt","fallbacks": ["rednote/dots-ocr"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

GPT alternativesdots.ocr alternativesBest OCR models in 2026