For $0.013 a page, GPT-4o reads your PDFs so you don't have to

What it does gptpdf converts PDFs into Markdown by rendering each page, marking non-text regions with bounding boxes via PyMuPDF, and asking a multimodal LLM like GPT-4o to transcribe the annotated result. The tool extracts both the structured text and any embedded images, returning a Markdown file plus a list of image paths. The authors claim it handles typography, math formulas, tables, pictures, and charts, and cite an average cost of about $0.013 per page.

The interesting bit The whole pipeline is just 293 lines of code. Rather than building a complex document engine, the project treats PDF parsing as a computer-vision task: point the model at the tricky bits and let it infer the layout.

Key highlights

• Exposes a single function parse_pdf with tunable prompts (prompt, rect_prompt, role_prompt) to guide how the model interprets text and marked regions.
• Supports any OpenAI-compatible vision endpoint, including GPT-4o, GLM-4V, Qwen-VL-Max, Yi-Vision, and Azure OpenAI.
• Returns extracted images alongside Markdown output, keeping figures and diagrams separate from the text stream.
• Ships with example conversions of real academic papers, such as Attention Is All You Need.

Caveats

• Requires an external vision-LLM API; the README does not mention support for fully local or offline execution.
• The “almost perfect” parsing claim is qualitative; no benchmarks or error rates are provided against traditional extractors.

Verdict Useful if you need readable Markdown from complex, layout-heavy PDFs and can justify per-page API spend. Look elsewhere if you need deterministic, offline extraction.