The 3 causes of "something went wrong while analyzing your file"

The wording is vague because the failure happens inside a sandbox you cannot see. It still tells you something useful: where in the pipeline it broke.

What the message actually means

Attaching a file starts two separate things. First the file transfers to OpenAI. Then a tool opens it, either text extraction for documents or a Python sandbox for spreadsheets and data files. This message is the second stage failing.

That rules a few things out straight away. It is not the 512 MB ceiling, because size is checked on the way in and produces a size error instead. It is not usually an unsupported file type either, since those are refused rather than analysed. Your file got in. It just could not be read once it was there.

Work through it in this order

1. Retry once

   Transient failures are real and one retry costs you almost nothing. One.

2. Try again in a fresh chat

   Attaching the same file to a brand new conversation sometimes works when a long one keeps failing.

3. Upload a tiny test file

   A two line CSV, or a text file with one sentence in it. If that fails too, your document is innocent. Check status.openai.com and wait.

4. Open the file on your own machine

   Not a preview pane. Open it properly. Does it ask for a password? Is it zero bytes? Does the application complain about the format?

5. Re-export it to something plain

   CSV instead of XLSX, plain text instead of DOCX. If a plain export of the same content works, the original file was the problem.

Cause one: the file cannot be opened at all

This one is deterministic. It fails at the same point every time, usually quickly.

It is protected. A PDF with an open password cannot be read. Less obviously, permissions passwords that only restrict copying and printing can also block text extraction. Open the file, save an unprotected copy, upload that.

It is truncated or corrupt. Interrupted downloads, files dragged out of an email preview, and exports that failed halfway through all produce something that looks like a file and is not one.

The extension is lying. An old .xls renamed to .xlsx, a .doc renamed to .docx, or an HTML table that a reporting tool saved with a spreadsheet extension. The upload trusts the name. The parser does not. Open it in the real application and save it again in the format it claims to be.

It is a cloud placeholder. iCloud Drive, OneDrive and Dropbox all show files that are not actually on your disk. Make sure the file is downloaded locally before you attach it.

It is empty. Check the size in Finder or Explorer. Zero byte exports happen more often than people expect.

Cause two: it opens, then processing gives up

Different signature. It thinks for a while first, sometimes tens of seconds, and it can fail at a different point on each attempt. Occasionally it half works and then breaks partway through an answer.

The published ceilings are about 50 MB for spreadsheets and 2 million tokens for text and document files. Under those numbers a file can still be too much work. OpenAI does not publish a time or memory budget for the analysis sandbox, so treat this as observed behaviour rather than a stated rule.

The files that trigger it tend to share traits: spreadsheets with hundreds of thousands of rows, workbooks stuffed with formulas, pivot tables and multiple sheets, CSVs with hundreds of columns, PDFs running to thousands of pages, and anything with an unusual text encoding.

The fix is to give it less work. Export the one sheet you care about as CSV. Delete the columns you will not ask about. Split by date range or by section. Converting XLSX to CSV drops formulas, formatting and charts, which is exactly the weight that was causing trouble.

Telling the three apart

If the tiny test file uploads and analyses fine, stop blaming OpenAI. If it does not, stop editing your document.

The route that skips the analysis step

For reading work rather than data work, there is a way around all of this. Pasting text into the chat is not a file upload, so there is no file for a parser to choke on and no sandbox to time out. The content arrives as ordinary messages.

That will not help you if you need charts, pivot tables or code run over a dataset, because those need the file. It helps a great deal when the file is a long report, a contract or a book and all you wanted was for ChatGPT to read it.

If none of it works

Report it, but only once you have the evidence. A tiny file that analyses correctly, a clear status page, a document that opens fine on your machine, and a plain text or CSV export that still fails together make a real bug report. OpenAI asks for your account email, a screenshot, the timestamp with your timezone, your platform and browser, and the request ID if you have one.

Until then the practical decision is simple. If you need calculations run over the file, shrink it or split it until the analysis step copes. If you only need the content read, stop uploading and send the text instead.