OCRRouter

GPT vs SmolDocling

Pick GPT when you need handwriting, custom JSON extraction or document question answering through a hosted API. Pick SmolDocling only for English page-conversion experiments on a small budget where you can self-host, because it is a preview model that its authors say is superseded by Granite-Docling. GPT costs per token and gives no coordinates. SmolDocling is free to run and returns bounding boxes.

GPTopenai/gptSmolDoclinghuggingface/smoldocling
TypeVision modelOpen source
Price11 credits / page1 credit / page
Per 1,000 pages$11.00$1.00
Typical speed2.4 s / page1.0 s / page
Best forMessy scans, custom instructionsVery cheap page conversion
InputsPDF, imagesPDF, images
Languages50+ languagesMainly English
Max pages per request100500
Tables✓ Yes✓ Yes
Handwriting✓ YesNo
Math✓ Yes✓ Yes
Bounding boxesNo✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ YesNo
On OCRRouterAvailableAvailable

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 is reported to handle messy cursive well: one HN user saw about 5% errors versus about 30% for Mistral OCR.
  • You can prompt it to extract fields into custom JSON and to answer questions about a document, not just transcribe it.
  • It runs as a hosted API that takes files up to 50 MB per request, so there is no GPU to set up.

Choose SmolDocling if

  • It is free to self-host, so you pay only for compute, and at 256M parameters it runs on modest hardware.
  • It returns bounding boxes and location tags, plus DocTags, Markdown, HTML, JSON and LaTeX, which GPT's standard response does not give.
  • It runs at about 0.35 seconds per page on an A100 with vLLM and handles text, tables, code and formulas in one model.

Key differences

GPTSmolDocling
How you run itHosted OpenAI APISelf-hosted, 256M parameters
LicenceProprietaryApache-2.0 per the model summary (the metadata tag says CDLA-Permissive-2.0)
PriceToken-based; page images add input tokens, so cost per page is not fixedFree, you pay for compute
CoordinatesNo bounding boxes or confidence scores in the standard responseBounding boxes and location tags in the output
Output formatsMarkdown, JSON, plain textDocTags, Markdown, HTML, JSON, LaTeX, bounding boxes
BenchmarkOmniDocBench v1.5 overall 85.50 for GPT-5.2 (independent)Full-page OCR F1 0.80 and table TEDS 0.82 on FinTabNet, reported by the authors
Known failureCost per page varies with page imagesOutput can loop with endless token repetition or missing closing tags; English only

Sources

What each one is

GPT

General vision model. Good at messy scans and following extraction instructions.

GPT details

SmolDocling

Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.

SmolDocling details

Try 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": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

GPT alternativesSmolDocling alternativesBest OCR models in 2026