Your PDF is “too large” for ChatGPT — which of the three walls did you hit?

Wall one: megabytes — the one you probably didn't hit

Per OpenAI's File Uploads FAQ, the hard cap is 512 MB per file. Almost no text-based PDF gets near it — a 512 MB PDF is usually a scan archive or a brochure stuffed with print-resolution images. If your file genuinely exceeds the megabyte cap, a PDF compressor will get it through the door. But be clear about what that buys you: compression shrinks the images and fonts inside the container, not the amount of text the model has to read — so if the real problem is one of the next two walls, you've fixed nothing.

Wall two: tokens — the cap most “too large” PDFs actually trip

The same FAQ caps any text or document file at 2 million tokens. Sounds enormous, but PDFs spend tokens faster than you'd think: AI chats process PDF pages as images as well as text so they can read charts and layout, which multiplies the cost several times over — the per-page arithmetic is here. A long technical manual, a book scan with a text layer, or a report with hundreds of pages can trip the token cap while sitting at a perfectly innocent file size. That's the telltale, in fact: a small-looking file rejected as too large is a token problem, not a megabyte problem.

Wall three: attention — the upload works, the answers don't

The sneakiest version has no error at all. The file passes both caps, but answers only reflect the beginning of the document, quotes come back garbled, or the model plainly ignores later chapters. That's the working context filling up: what the model can attend to in a conversation is far smaller than what the upload meter accepts. Splitting the PDF into five uploads doesn't help — you've turned one overflowing conversation into five, and burned your upload allowance doing it.

The fix, in order of preference

• Convert only the pages you need. Most “too large” PDFs are really “I need chapter 3 of something huge”. The PDF to Markdown converter has a page-range field — type “40-65” and you get just those pages as clean Markdown, with the original page numbers kept so citations still line up. Paste that into the chat: no upload, a tiny fraction of the tokens, and the model's attention spent only on what matters.
• Convert the whole document if you truly need all of it. Markdown strips the image-processing overhead and the layout baggage; the same content typically lands at a small fraction of the PDF's token cost. The token counter on the result card tells you before you paste whether it fits comfortably.
• Split the file only as a last resort — it multiplies uploads (which are metered) and fragments the model's view of the document. If you must, split by chapter, not by equal page counts, so each piece is self-contained.