ChatGPT file upload limit: which one stops you

ChatGPT rejects a file when it breaks one of four published caps: size, tokens, upload rate or storage. Two other limits never produce an error. A file can be accepted and still be answered from only part of its text, and pictures inside some file types are dropped without a word.

A rejected PDF and a half-read PDF need different fixes. Telling them apart takes a few minutes, while guessing on the free plan costs part of a very small daily allowance.

File size and upload limits OpenAI publishes, by plan

A paid subscription raises how often files can go in, not how large any one file can be. Every plan shares the same per-file ceiling, and text or document files carry a second cap counted in tokens. A token is a small piece of text, roughly three quarters of a word, which is why a short-looking file can still be too long.

Spreadsheets skip the token cap and are limited by size instead, depending on how heavy each row is. Images have their own cap per picture. Storage is a separate pool shared by chats, Projects and custom GPT knowledge, so it fills up from everything uploaded over time.

The rate limit is the row where plans differ. Paid plans may upload up to 80 files every 3 hours(as of October 2026), and the free plan is limited to 3 uploads a day. OpenAI adds that it may lower these limits during peak hours, so the “Published ChatGPT file caps, as of October 2026 (sizes in MB or GB)” table shows the best case.

Published ChatGPT file caps, as of October 2026 (sizes in MB or GB)

The size caps are the same on every plan. Only the upload-rate rows change, from 3 files a day on the free plan to 80 files every 3 hours on paid plans.

Limit	Value	What it counts
File size	512MB	One file of any type
Document length	2M tokens	Text and document files, not spreadsheets
Spreadsheet size	about 50MB	CSV and spreadsheet files, varies by row
Image size	20MB	One image
Rate, paid	80 files per 3 hours	Rolling window, may be lowered at peak
Rate, free	3 files per day	All file types together
Storage, user	25GB	Chats, Projects and GPT knowledge
Storage, organization	100GB	Every user in the organization

Accepted is not read: the Enterprise 110k-token cut

A file that clears every published cap can still be answered from only part of its text. The caps in the “Published ChatGPT file caps, as of October 2026 (sizes in MB or GB)” table decide whether ChatGPT takes the file. They do not decide how much of it the model holds in mind when it replies, and that second budget is much smaller.

ChatGPT Enterprise shows the difference clearly because OpenAI publishes its numbers. The models have a context window, the amount of text they can consider at once, of 128k tokens, roughly 200 pages of text. Up to 110k tokens of that window can come from uploaded documents.

For a single document longer than that, only the first 110k tokens are included, starting from the beginning. The remainder goes only to the private search index. OpenAI's testing suggests about 350 pages of mixed text and images fill the window.

Page count is a poor guide for this reason. Two documents with the same page count can behave very differently, because tokens rather than pages decide how much is read. A question about the last chapter of a long report is answered from search results rather than from the text in front of the model, so a long report is safer split into parts.

ChatGPT does not show how much upload quota is left

The rolling count is hidden from users, and the one figure that can be looked up measures something else. OpenAI states that ChatGPT does not currently show how much of the rolling upload-rate quota has been used or how much remains. The 80-files-per-window figure is therefore a limit that cannot be tracked on screen.

Two kinds of limits exist. The rolling upload rate is one, and the shared storage caps that apply across chats, Projects and custom GPT knowledge are the other. Storage usage can be checked in the Storage setting, and deleting old files brings it down.

The rate behaves differently. Nothing on screen counts it, and an error appears once a user or organization cap has been hit. Deleting files does not undo uploads already made in the window.

Uploads are best spent on purpose. Combining related documents into one file before starting uses a single upload where several would otherwise be needed, which matters most on the free plan.

A failed upload can still count against your rate cap

Retrying a rejected file is risky because the rejected attempt may already have been charged. OpenAI states that some failed upload attempts count toward the upload-rate cap. Since the remaining quota is invisible, there is no way to see whether a failure was billed.

On the free plan the effect is easy to measure. With 3 uploads a day, one retry can spend 1 of the 3. Two failures and one success can use the whole day, and nothing on screen reports which attempts were counted.

The sensible order is to check the file before the second attempt. A document over the size cap will fail again, and so will a spreadsheet over its own cap. If either number is close, split the file first and send the parts.

Convert to PDF first: images in other files are skipped

Pictures embedded in a Word or PowerPoint document never reach the answer unless the document is exported to the format OpenAI recommends. ChatGPT Enterprise processes files in three ways: text extraction, code analysis and image interpretation. Images embedded in files other than PDFs are not processed, so a chart pasted into a Word report or a PowerPoint deck adds nothing to the reply.

OpenAI's instruction is to convert the file to a PDF before uploading. The saved PDF carries the pictures along, and that one step often separates a summary that mentions the figures from one that describes only the paragraphs around them.

Whether the pictures are then read depends on the plan. Reading visuals such as images, graphs and diagrams inside a PDF is available only to ChatGPT Enterprise, and it is not supported for Free, Pro, Team or Edu accounts. On those plans a PDF of a chart is read as the text near the chart.

Where the PDF is attached also matters. A PDF added as GPT Knowledge or as a Project File uses text-only retrieval, while the same PDF uploaded in a conversation can use visual retrieval. Spreadsheets follow a separate rule: ChatGPT Enterprise always uses Code Interpreter for them, even when the document contains a lot of text.

What to do when you hit a limit: fix, split or paste

A size problem is solved by making the document smaller, and a count problem by sending text instead of files. The error message shows which cap was hit. A user or organization storage cap is named in the message, while a rate problem appears as an upload limit message.

For a file over the size cap, shrink it or split it. Email Attachment Size Limits: Fit Your PDF covers getting a PDF under a fixed size, and splitting a long report also helps with the reading budget.

When the rate cap is the problem and the content is plain text, pasting it into the message body sends no file at all. That uses none of the upload allowance.

Extensions need a look as well. Most common formats go through, including XLSX, XLS, CSV, TSV, DOCX, PPTX, PDF and TXT, but .gdoc is not supported.

Summary: which limit stopped you and what to do next

Four published caps announce themselves with errors, while the reading budget and the skipped pictures give no warning at all. The published numbers cover acceptance only.

On ChatGPT Enterprise a document over 110k tokens contributes its first 110k tokens to the context window, and images outside a PDF are skipped on every plan. Which fix fits depends on what went wrong, and each case below has one next step.

• An error on a large file: shrink or split it, and check the size before any retry.
• An error on the free plan after a failed attempt: wait for the daily allowance, because the failure may have been counted.
• An answer that ignores the end of a long report: split the report into parts, since tokens rather than pages set the reading budget.
• An answer that ignores charts in a Word or PowerPoint file: save the file as a PDF and upload that.
• Text content with the rate cap reached: paste it into the message.