GPT vs Surya
Pick GPT for handwritten notes, custom JSON extraction and document question answering, where you accept a per-token bill and no coordinates. Pick Surya for local, free OCR and layout work that needs bounding boxes and reading order across 90+ languages. The main split is a hosted proprietary API versus a self-hosted open toolkit.
GPTopenai/gpt | Suryadatalab/surya | |
|---|---|---|
| Type | Vision model | Open source |
| Price | 11 credits / page | 2 credits / page |
| Per 1,000 pages | $11.00 | $2.00 |
| Typical speed | 2.4 s / page | 1.5 s / page |
| Best for | Messy scans, custom instructions | Layout and line-level OCR |
| Inputs | PDF, images | PDF, images |
| Languages | 50+ languages | 90+ languages |
| Max pages per request | 100 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | No | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | ✓ Yes |
| On OCRRouter | Available | Available |
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 was reported to make about 5% errors on messy cursive, against about 30% for Mistral OCR, so it suits handwritten notes.
- You can prompt it to extract data into custom JSON and ask questions about a document, which a layout and OCR toolkit does not do.
- It is a hosted API that takes files up to 50 MB per request, so you run no inference server.
Choose Surya if
- It is free to self-host under Apache-2.0 code, so you pay only for compute, as long as your funding or revenue stays under $5M.
- One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes.
- At 689M parameters it runs on a laptop CPU or Apple Silicon via llama.cpp, with peak RAM of about 4-6 GB on an M4 Max.
Key differences
| GPT | Surya | |
|---|---|---|
| How you run it | Hosted OpenAI API | Self-host: vLLM server on an NVIDIA GPU, or llama.cpp on CPU or Apple Silicon |
| Licence | Proprietary | Code Apache-2.0; weights under a modified AI Pubs Open Rail-M, free for research, personal use and companies under $5M funding or revenue |
| Price | Per token. PDF page images add input tokens, so cost per page is not fixed. | Free self-hosted. A commercial licence is needed above $5M revenue or funding. |
| Output formats | Markdown, JSON, plain text | JSON, HTML, LaTeX, bounding boxes, plain text |
| Coordinates | No bounding boxes or confidence scores in the standard response | Line and block level bounding boxes |
| Benchmark | GPT-5.2 scored 85.50 on OmniDocBench v1.5 and 86.59 on v1.6, run by independent parties | Surya 2.5 scored 83.7% on olmOCR-bench and 85.9 on OmniParseBench, reported by its authors |
| Known weak spots | Cost per page varies with page images; GPT-5.2 trailed Gemini 3 Pro (90.33) on OmniDocBench v1.5 | Handwriting, complex multilingual tables (58.1) and photos or natural scenes; Surya 2.5 table output lacks row and column geometry |
Sources
- OpenAI file inputs docs
- OmniDocBench leaderboard
- HN Mistral OCR thread
- Surya GitHub README
- Datalab pricing and license FAQ
What each one is
Surya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya detailsTry 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": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}