GPT vs Donut
Pick GPT when you need a working result today: handwritten notes, custom JSON extraction with reasoning, or questions over a document, all through a hosted API with no training. Pick Donut when your documents are a narrow, fixed layout such as receipts or forms and you want an MIT-licensed model you fine-tune and host yourself. The main split is a general hosted model against a base model that needs training before it does anything useful.
GPTopenai/gpt | Donutnaver/donut | |
|---|---|---|
| Type | Vision model | Open source |
| Price | 11 credits / page | 1 credit / page |
| Per 1,000 pages | $11.00 | $1.00 |
| Typical speed | 2.4 s / page | 1.0 s / page |
| Best for | Messy scans, custom instructions | Receipt and form fields after fine-tuning |
| Inputs | PDF, images | images |
| Languages | 50+ languages | English, Chinese, Japanese, Korean |
| Max pages per request | 100 | 100 |
| Tables | ✓ Yes | No |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | No | ✓ Yes |
| Multilingual | ✓ Yes | No |
| 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 works out of the box through the OpenAI API, and you can prompt it to transcribe or to extract fields into custom JSON.
- It suits handwriting. An HN user reported about 5% errors on messy cursive, versus about 30% for Mistral OCR.
- It can output Markdown, JSON or plain text, and answer questions about a document, with files up to 50 MB per request.
Choose Donut if
- It is MIT licensed and free to self-host, so you pay only for compute and keep documents on your own machines.
- It maps document images straight to structured output with no separate OCR step, which fits receipt and form field extraction after fine-tuning.
- A practitioner on HN says about 200 documents of fine-tuning can reach human-level accuracy on narrow data.
Key differences
| GPT | Donut | |
|---|---|---|
| How you run it | Hosted API from OpenAI | Self-hosted; load it with Hugging Face Transformers |
| Licence | Proprietary | MIT |
| Price | Token-based per model; page images add input tokens, so cost per page is not fixed | Free; you pay for compute |
| Ready to use | Yes, prompt it to transcribe or extract | No. The base checkpoint is pretrain-only and must be fine-tuned |
| Output formats | Markdown, JSON, plain text | JSON, plain text |
| Benchmark | GPT-5.2 scored 85.50 on OmniDocBench v1.5 and 86.59 on v1.6, versus 90.33 for Gemini 3 Pro on v1.5 | Not published |
| Known failure | No bounding boxes or confidence scores in the standard response | Misses tiny text in large images because of input resolution limits; Transformers v5 dropped image-to-text pipeline support, so use direct loading or v4 |
Sources
- OpenAI file inputs docs
- OmniDocBench leaderboard
- HN Mistral OCR thread
- donut-base model card
- Donut ECCV paper
- Hacker News thread
What each one is
Donut
OCR-free document understanding model. Fine-tuned for receipts, forms and document Q&A.
Donut 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": ["naver/donut"],"file_url": "https://example.com/document.pdf"}