CSV and Excel limits in ChatGPT: about 50 MB, and no token cap

The exemption is the important part

OpenAI's documentation states the token cap on text and document files, then adds directly that this limitation does not apply to spreadsheets.

That single sentence explains a lot of otherwise confusing behaviour. A 300 page report gets truncated silently. A CSV with 200,000 rows goes through. Both are enormous amounts of text, and they are governed by different rules.

So for spreadsheets, size is genuinely the limit. For documents it almost never is.

CSV beats XLSX nearly every time

An XLSX file carries far more than its data: cell formatting, styles, fonts, conditional formatting rules, embedded charts, images, and sometimes cached pivot table data. None of that is what you want analysed.

Exporting to CSV strips all of it and leaves the numbers. Files routinely drop by an order of magnitude, and nothing ChatGPT would have used is lost.

1. Open the workbook

   Excel, Numbers, LibreOffice, Google Sheets.

2. Save or export each sheet as CSV

   CSV holds one sheet, so a multi sheet workbook becomes several files. Only export the sheets that matter.

3. Upload the CSV

   Smaller, faster, and no formatting to get in the way.

Making the data readable

Spreadsheets built for humans often confuse analysis. A few things that consistently help:

One header row, at the top. Merged cells, multi row headers and a title row above the headers all make it harder to work out what a column is.

No blank rows in the middle. They read as the end of the data.

Consistent columns. A column with numbers in most rows and text in a few is a common source of wrong answers, because it is unclear what the column is.

Say what the columns mean. Especially abbreviations. q3_rev_gbp_net is obvious to you and guessable at best to anything else.

If the file is still too large

Send fewer columns. Most analyses need a handful. Delete the rest before uploading.

Send a sample. For "what patterns are in this", a few thousand representative rows produce nearly the same answer as two hundred thousand.

Split by period or category. Twelve monthly files often work better than one annual one, because each question is asked against the relevant slice.

Aggregate first. If your question is about totals by region, upload the totals by region rather than every transaction.

Working with the data once it is in

Ask for the analysis rather than the summary. "Summarise this spreadsheet" produces a description of the columns. This produces something useful:

Group by region, sum revenue, and tell me which regions grew and which shrank against the previous quarter. Show the numbers you used.

Requiring the numbers matters. It lets you check the arithmetic, and arithmetic over a large dataset is exactly where confident mistakes happen.