How many PDFs can you upload, and how many get read?

People ask this about PDFs in particular because PDFs are what they have a lot of: a folder of research papers, a stack of invoices, four versions of the same contract. The count answer is mildly boring. ChatGPT treats a PDF exactly like a text file or a spreadsheet, one slot each, regardless of what is inside it.

The useful answer is the second half. Uploading ten PDFs and having ten PDFs read are different achievements, and the limits that decide the second one are not counts at all.

The counts, by plan

These stack rather than replacing each other. Having daily allowance left does not help inside a project that is already full, and an empty project does not help if your storage is at the ceiling.

Nothing on that table is about PDFs

There is no published PDF count anywhere, and that is worth understanding rather than just noting. A three page PDF costs one slot. A nine hundred page PDF costs one slot. The caps count files.

Which cuts both ways. Merging six short PDFs into one before uploading turns six slots into one, and on Free that is the difference between two days of allowance and part of an afternoon. But an enormous PDF gets no extra room for being large, and it is the one most likely to be truncated.

The limit you actually meet first

Text and document files are capped at 2 million tokens. That is the ceiling that decides how much of a PDF gets read, and it is nothing to do with how many you sent.

The failure mode is what makes it awkward. The PDF is not rejected. It uploads normally, ChatGPT reads whatever fitted, and answers with total confidence about a document it has seen part of. No warning, no marker in the reply.

Size is the limit people expect and the one that almost never bites. The ceiling is 512 MB, and a text-based PDF of several hundred pages is usually a few megabytes.

Scanned PDFs spend a slot and return nothing

Every plan except Enterprise handles documents as text only, except Enterprise. So a scanned PDF, which is pictures of pages rather than digital text, uploads successfully, counts against your allowance, and yields no text to read.

Three of those on Free is your entire day for nothing. Check before you spend the slot: open the PDF and try to select a sentence. If the cursor draws a box instead of highlighting words, it is a scan and it needs OCR first.

Projects and custom GPTs hold their files

The project caps are not per conversation. Files you added to a project in March are still sitting in it, still counted, until you remove them. When a project refuses a new PDF, look at what is already in there rather than at your daily allowance.

Custom GPTs are stricter and catch people out badly. 10 files per GPT, for the lifetime of that GPT, not per chat and not per month. If your GPT needs to know twenty documents, combine them before you start rather than finding the ceiling at document eleven.

Underneath everything sits 25 GB, shared across chats, projects and GPT knowledge. PDFs are among the heaviest things people upload, so this fills faster with PDFs than with anything else. Settings, then Storage, shows you where it stands. It shows storage only, so there is no display anywhere for how much of the rolling upload rate is left.

Stretching the allowance you have

Merge before uploading. Any PDF tool will combine several files into one. The caps count files, so ten short reports become one slot.

Send the pages you need. Extracting chapter four into its own PDF costs the same one slot and usually produces a better answer, because there is less irrelevant material competing for attention.

Convert to text. A .txt export of a PDF is a fraction of the size and drops the images that were going to be discarded anyway on every plan except Enterprise. It does nothing for length, but it removes size and image handling from the equation.

Paste instead of uploading. Pasted text is a message, not an upload, so it touches none of the counts above. When you are completely out of uploads, this still works.

What to check before you upload

The count is the easiest limit to reason about and usually the least relevant. Before you spend a slot, two questions settle most of it.

Does the PDF contain selectable text, or is it a scan that will return nothing? And is it short enough to fit inside 2 million tokens, or should you send the section you actually care about instead of the whole thing?

Get those right and you will rarely notice the counts. Get them wrong and you can burn a day of allowance on files ChatGPT never really read.