Stop feeding PDFs to ChatGPT and Claude. Convert them first.

Uploading a PDF to an AI model costs several times more tokens than the same content as plain text — and often gives worse answers. Here's why, and what to do instead.

If you paste a long PDF into ChatGPT or Claude and watch your usage disappear faster than you expected, you're not imagining it.

Why a PDF costs so much more

A PDF isn't text as far as a model is concerned. It's a layout format — text, fonts, positions and images all mixed together. To read one reliably, these tools generally process each page as an image as well as pulling the text out.

Images are expensive. A single page rendered as an image typically costs well over a thousand tokens. The same page as plain text is usually a few hundred. Across a fifty-page report, that difference is enormous.

It's not just cost — it's often accuracy

PDF layout carries no meaning the model can rely on. A heading is just larger text at a certain position; a table is lines and coordinates. Converting to Markdown first makes that structure explicit — a heading becomes a heading, a list becomes a list — and models handle explicit structure noticeably better.

Where this matters most

• Long documents — reports, research papers, manuals, contracts
• Anything you'll ask about repeatedly, since you pay the cost every time
• Building with the API, where tokens are billed directly
• Feeding documents into a knowledge base, Notion or Obsidian

What to do instead

Convert the PDF to Markdown or plain text once, then work from that. It's smaller, cleaner, easier for the model to follow, and you can read it yourself to check nothing important was mangled before you rely on it.

Convert it first

This only works on PDFs that contain real text. If your file is a scan, there are no words in it to extract — only a picture of words. You'd need OCR for that.

One thing to check first

Open your PDF and try to select a sentence. If you can highlight it, the text is really there and conversion will work. If your cursor just draws a box, it's a scan.

Common questions

Why not just upload the PDF to ChatGPT or Claude?

You can, and for a one-off question it is fine. The cost shows up when you ask repeatedly: every question re-sends the whole document, including page images carrying no information you are asking about. Markdown sends the words and nothing else.

Does converting to Markdown lose anything?

Layout, yes — columns, exact positioning, page breaks. Meaning, mostly no. Headings, lists and paragraph structure survive, and those are what a model actually uses. Tables are the exception worth checking, since a complicated one can come out flattened.

Will the model understand Markdown headings?

Yes. Markdown is close to the format these models were trained on, so a document with real headings and lists is easier for them to navigate than the same text extracted flat.

What if my PDF is a scan?

Then there is no text to extract — it is a picture of words. Run it through Searchable PDF first, which uses OCR to add a real text layer, then convert.