ChatGPT PDF page limit: none published, but 2M tokens caps it

Why pages are the wrong unit

A page is not a fixed amount of anything. A page of a legal contract holds several times the text of a page in a report with a chart on it. So "how many pages" has no stable answer, and OpenAI does not publish one.

What exist are three real ceilings.

Length is the one that gets you

Almost every "it stopped reading my PDF" is this.

Text and document files are capped at 2 million tokens. As a rough feel, that is a very large amount of pure text, but a dense document reaches it far sooner than people expect, and it is reached long before the size ceiling on anything text-heavy.

What makes it awkward is the failure mode. The document is not rejected. It uploads, ChatGPT reads what fitted, and answers confidently from that. There is no warning, no marker, and no hint in the response that anything was left out.

Where the number 100 comes from

People search for a page limit partly because of a message about adding at most 100 images, including PDF pages, to a chat.

That is an image ceiling rather than a reading limit, and it applies per conversation. PDF pages count toward it when a PDF is handled visually, which is why one long document can trip it.

Two things follow. Starting a new chat resets it, since it is per conversation. And converting the PDF to text removes it from the equation completely, because text is not images.

Getting a long PDF read in full

In order of effort.

Send the section you need. Extract those pages into their own PDF. Usually a better answer too, since there is less irrelevant material competing for attention.

Convert to text. Strips page geometry and embedded images, leaving the words. This helps a lot with size and nothing at all with length, so it depends which ceiling you hit.

Split it and send the parts in order. Tell ChatGPT up front that several parts are coming and not to answer until the last one arrives. Skip that and it summarises part one as though it were the whole document.

Use a different tool at book length. Past a certain size, even parts sent in order fall out of memory before the last one lands. A retrieval tool that searches the document and surfaces relevant passages will beat chunking every time.

If pages come back empty rather than partial

Different problem entirely. If ChatGPT reports finding nothing at all, your PDF is a scan: images of pages with no digital text in them. Every plan except Enterprise discards images, so there is nothing to extract.

Test it by trying to select a sentence in the PDF. If your cursor draws a box instead of highlighting words, it is a scan, and it needs OCR before any of this page applies.