GPT vs Qwen-VL
Pick GPT for handwritten notes, custom JSON extraction and document question answering, where you want a hosted API and no GPU to run. Pick Qwen-VL when you need open weights under Apache-2.0, layout positions, or self-hosting, and add loop checks on low-quality scans. The main split is deployment: GPT is a proprietary API, Qwen-VL is a model you can run yourself.
GPTopenai/gpt | Qwen-VLqwen/qwen-vl | |
|---|---|---|
| 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 | 2.0 s / page |
| Best for | Messy scans, custom instructions | Flexible extraction with prompts |
| Inputs | PDF, images | PDF, images |
| Languages | 50+ languages | 30+ languages |
| Max pages per request | 100 | 200 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | ✓ Yes |
| Math | ✓ Yes | ✓ Yes |
| 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
- An HN user reported about 5% errors on messy cursive versus about 30% for Mistral OCR, so it suits handwritten notes.
- It extracts into custom JSON with reasoning and fits document question answering, with files up to 50 MB per request.
- PDF input sends both extracted text and page images to the model, so you get one hosted call with no GPU setup.
Choose Qwen-VL if
- Apache-2.0 open weights in sizes from 2B to 235B-A22B MoE let you self-host and pay only for compute.
- It returns bounding boxes and a Qwen HTML document format, which helps when you need layout positions.
- OCR covers 32 languages, and it handles low light, blur and tilted text per Qwen, so it fits photos and screenshots with text.
Key differences
| GPT | Qwen-VL | |
|---|---|---|
| Deployment | Hosted API only | Self-hosted open weights, or hosted on Alibaba Cloud |
| Licence | Proprietary | Apache-2.0 (Qwen3-VL repository) |
| Price | Token-based per model; page images add input tokens, so cost per page is not fixed | Free self-hosted, you pay for compute; Alibaba Cloud hosted pricing varies by model |
| Output formats | Markdown, JSON, plain text; no bounding boxes or confidence scores in the standard response | Plain text, HTML, bounding boxes, JSON, Markdown |
| Benchmark | GPT-5.2: 85.50 on OmniDocBench v1.5, 86.59 on v1.6 | Qwen 3 VL 8B: 64.6 on olmOCR-bench, below dedicated OCR models |
| Known failure | Cost per page is not fixed and there are no word coordinates, so it is a poor fit for the cheapest bulk OCR | Can loop repeating the same text, worse with greedy decoding and on low-quality scans |
| Hardware | None to manage | FP8 checkpoints need H100-class GPU and CUDA 12+; the 235B example uses 8-way tensor parallel on vLLM |
Sources
- OpenAI file inputs docs
- OmniDocBench leaderboard
- HN Mistral OCR thread
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
What each one is
Qwen-VL
Alibaba's open-weight vision language models (Qwen2.5-VL, Qwen3-VL). Strong general OCR and grounding.
Qwen-VL 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": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}