GPT vs Granite Docling
Pick GPT when you need handwriting, custom JSON extraction with reasoning, or document question answering, and you can accept a hosted API with variable per-page cost. Pick Granite Docling for English PDFs with tables, code and equations that you want to convert cheaply on your own CPU or small GPU, with bounding boxes in the output. The main split is hosted and proprietary versus a 258M self-hosted Apache-2.0 model with layout-aware output.
GPTopenai/gpt | Granite Doclingibm/granite-docling | |
|---|---|---|
| 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 | 1.5 s / page |
| Best for | Messy scans, custom instructions | Small-model document conversion |
| Inputs | PDF, images | PDF, images |
| Languages | 50+ languages | English, experimental CJK and Arabic |
| Max pages per request | 100 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | No | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| 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 handles handwritten notes: an HN user reported about 5% errors on messy cursive versus about 30% for Mistral OCR.
- It can extract into custom JSON and answer questions about a document, which Granite Docling is not listed for.
- It takes files up to 50 MB per request through the API, so there is nothing to host or tune.
Choose Granite Docling if
- It is Apache-2.0 and free to self-host, so you pay only for compute, and at 258M parameters it runs on CPU or a small GPU.
- Its authors report table structure TEDS of 0.97 on FinTabNet, code recognition F1 of 0.988 and equation F1 of 0.968.
- It outputs DocTags, Markdown, HTML, LaTeX and bounding boxes, which suits pipelines that need coordinates.
Key differences
| GPT | Granite Docling | |
|---|---|---|
| How you run it | Hosted OpenAI API | Self-hosted, typically inside Docling |
| Licence | Proprietary | Apache-2.0 |
| Price | Token-based; page images add input tokens, so cost per page is not fixed | Free; you pay for compute |
| Output formats | Markdown, JSON, plain text; no bounding boxes or confidence scores in the standard response | DocTags, Markdown, HTML, LaTeX, bounding boxes |
| Benchmark | GPT-5.2 scored 85.50 on OmniDocBench v1.5 and 86.59 on v1.6; Gemini 3 Pro scored 90.33 on v1.5 in the same paper | Table TEDS 0.97 (FinTabNet), full-page OCR F1 0.84, OCRBench 500, all reported by the authors |
| Languages | Not published | English primary; Japanese, Arabic and Chinese are experimental |
| Known failure | Cost varies with page images; no word coordinates | Repetitive loops on page headings and invoices; on GPUs without bfloat16 such as T4 it may output only exclamation marks unless run in float32 |
Sources
- OpenAI file inputs docs
- PaddleOCR-VL-1.5 paper (GPT-5.2 eval)
- HN Mistral OCR thread
- Granite-Docling model card
- HF discussion #24 repetition
What each one is
Granite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "openai/gpt","fallbacks": ["ibm/granite-docling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
GPT alternativesGranite Docling alternativesBest OCR models in 2026