Does the ChatGPT upload limit depend on the model?

Both things are true at the same time, and that is what makes the question confusing. Model choice genuinely affects how well a long document is handled. It affects none of the numbers that decide whether the document is accepted in the first place.

The limits are enforced before a model sees anything

An upload happens in two stages, and they are easy to blur together.

Stage one is acceptance. The file is measured against 512 MB per file, 2 million tokens for text and document files, about 50 MB for spreadsheets and 20 MB per image. Your recent activity is measured too, against 3 file uploads per day on Free and 80 files every 3 hours generally, and against your 25 GB of storage. Then the file is converted to text.

Stage two is reading. A model receives whatever text came out of stage one and answers your question about it.

The model picker sits in stage two. Every published upload limit belongs to stage one, and none of them is stated per model. That is the whole answer, and the rest of this page is about the part of the question that does have a real answer.

What actually varies with the model

The context window. How much of the conversation, including the extracted text of your file, the model holds in front of it at one time.

OpenAI publishes these on its pricing page, and the way it splits them is the part worth reading. The figures are given by plan and by whether you are using an instant or a reasoning model, not by the individual model name in the picker. On the instant models the total window is 27K tokens on Free, 54K tokens on Go, 54K tokens on Plus and 128K tokens on Pro. On the reasoning models it is 256K tokens on Go and Plus and 400K tokens on Pro, with the Free reasoning window given only as "varies".

So even the one thing that does move is moving with your plan as much as with your choice in the picker. A larger window keeps more of a long document available at once, which shows up as fewer answers that are confident about chapter one and vague about chapter nine.

OpenAI attaches a footnote to those numbers that matters more than the numbers themselves: the input share is smaller than the total window, because system instructions, tools, memory and internal processing are taking space out of the same window. That is the honest answer to why a document that should have fitted did not, and it is why no figure on this page will tell you how many pages you can count on.

One row out of nine. It is an important row, but it is not the row people are asking about when they go looking for a model that allows bigger uploads.

Why a bigger model does not un-truncate a file

The per-file token cap is a ceiling on the file, not a promise that any model reads all of it. 2 million tokens is far larger than the context window of the models you can pick, so on a genuinely long document two separate reductions are happening. The cap decides how much of the file survives upload. The window decides how much of the surviving text the model can consider at once.

Beyond the window, ChatGPT tends to search the document rather than read it end to end. That is observed behaviour rather than published policy, but it explains the pattern people report: precise, correct quotations from some passages, and blank spots elsewhere in the same file.

Three different walls that get mistaken for each other

Model message allowances. On Free and Plus, some models can be used only so much before you are moved to another one. This is a message limit and it has nothing to do with files. It is also the one limit in this list that involves model choice, and it points the opposite way to what people assume: the model changes because you ran out, not to give you more room.

Upload caps. 3 file uploads per day on Free, 80 files every 3 hours generally, counted per account. Failed attempts can count toward the rolling cap, which is why people reach it having successfully uploaded very little.

Storage. 25 GB per user, shared across chats, Projects and custom GPT knowledge. Full storage produces an upload error that looks like a rate limit and is not one.

Only the first has anything to do with the model picker, and none of the three is fixed by using it.

Switching model in the middle of a conversation

The file stays attached to the conversation, and the text was extracted once at upload, so a new model works from the same extracted text rather than re-parsing your original file. What does not carry over cleanly is whatever the previous model had built up in view.

1. Name the document again

   Say which attached file you mean rather than saying "the document". It removes the ambiguity in a thread that already contains several.

2. Ask for a quotation before analysis

   "Quote the first two sentences of section 4." A model that cannot produce them does not have that part in view, and you have learned that in one message instead of after a wrong summary.

3. Re-state conclusions you want kept

   Paste back the short summary the earlier model produced. Treat it as input to the new model, not as shared memory.

What to change instead of the model

Send less per file. Convert to plain text, drop appendices, references and boilerplate you will not ask about. A long report often loses a third of its length with nothing of value gone.

Split by section and ask per section. One chat per section keeps everything relevant inside the window, whichever model is answering.

Split and send in order when the document has to arrive whole, with an instruction to hold the answer until the final part lands.

Run OCR first if the pages are scans. No model in the picker reads an image inside a document, because the images were discarded during extraction, on every plan except Enterprise.

The practical decision

If your problem is a rejected upload, an exhausted allowance or a truncated file, the model picker is the wrong control and no setting in it will help. Change what you send.

If your problem is that the upload worked but the answers go thin over a long document, model choice is worth trying, alongside cutting the document down. Those are two different failures with two different fixes, and reaching for the model picker on the first one is the reason this question gets asked so often.