GPT vs olmOCR
Pick GPT when you need handwriting, custom JSON extraction or document question answering through a hosted API with no GPU to run. Pick olmOCR for large-scale PDF to Markdown, especially academic papers with equations, where self-hosting is under $200 per million pages. The main split is deployment: GPT is a pay-per-token API with variable page cost, olmOCR is an open 7B model you run on a 12 GB+ GPU or call through a hosted endpoint.
GPTopenai/gpt | olmOCRallenai/olmocr | |
|---|---|---|
| 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 | Bulk PDF to text for datasets |
| Inputs | PDF, images | PDF, images |
| Languages | 50+ languages | Mainly English |
| Max pages per request | 100 | 1,000 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | ✓ Yes |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | No | No |
| 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
- Handwritten notes: an HN user reported about 5% errors on messy cursive, versus about 30% for Mistral OCR.
- You want extraction into custom JSON with reasoning, or document question answering, in one prompt.
- You do not want to run a GPU: it is a hosted API that takes files up to 50 MB per request.
Choose olmOCR if
- Bulk conversion is cheaper: under $200 per million pages self-hosted, per Ai2, or $0.09 input / $0.19 output per million tokens on DeepInfra.
- The weights, training data and code are open, and the toolkit is Apache-2.0.
- It handles equations, tables and mixed printed and handwritten pages, and strips headers and footers in reading order.
Key differences
| GPT | olmOCR | |
|---|---|---|
| Deployment | Hosted API only | Self-host with vLLM on a GPU, or use third-party hosted endpoints |
| Licence | Proprietary | Apache-2.0 (toolkit); weights, training data and code released |
| Pricing | Token-based per model; PDF page images add input tokens, so cost per page is not fixed | Under $200 per million pages self-hosted; DeepInfra $0.09 input / $0.19 output per million tokens |
| Output formats | Markdown, JSON, plain text | Markdown, JSON |
| Benchmark | GPT-5.2: 85.50 on OmniDocBench v1.5, 86.59 on v1.6 (independent) | 82.4 on olmOCR-Bench v0.4.0 (published by the authors); not published on OmniDocBench |
| Hardware and limits | Files up to 50 MB per request; no hardware to manage | NVIDIA GPU with at least 12 GB VRAM and 30 GB free disk; out of memory reported on a 12 GB GPU (issue #424) |
| Known failure | No bounding boxes or confidence scores in the standard response | Unreliable image bounding box coordinates (issue #437); can get stuck on the last few pages (issue #433) |
Sources
- OpenAI file inputs docs
- OmniDocBench leaderboard
- HN Mistral OCR thread
- olmOCR GitHub README
- Issue #424 OOM
- Issue #437 bbox
What each one is
olmOCR
Open-source vision model from Ai2 (often searched as olm ocr) that turns PDFs into clean, reading-order text. Latest release: olmOCR 2.
olmOCR 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": ["allenai/olmocr"],"file_url": "https://example.com/document.pdf"}