Which AI assistant should I use to analyse a spreadsheet?

What do ChatGPT, Claude, and Gemini have in common?

More than the comparisons usually admit. Each will take a spreadsheet, work out what the columns mean, write and run code against the data, draw a chart, and then let you argue with the result in plain English. For open-ended exploration, where you do not yet know which question matters, all three are hard to beat.

They also share the thing that decides most real decisions: to read your file, each one has to receive it. The rows go to OpenAI, Anthropic, or Google, and the analysis happens on their infrastructure. Everything below is detail on top of that.

Where does ChatGPT fit?

ChatGPT is the strongest all-rounder of the three. It profiles a file quickly, writes Python to answer what you asked, and handles follow-up questions well. If you want one tool that does a bit of everything, this is usually it.

Its limits show up in three places. A spreadsheet of around fifty megabytes is the working ceiling, though row count matters more than megabytes — analysis stays comfortable into the low hundreds of thousands of rows and gets shaky beyond. The sandbox has a time budget, so a big job can stop partway with nothing to show. And the sandbox is temporary: come back tomorrow and the file, and the work built on it, are usually gone.

Where does Claude fit?

Claude is the one to reach for when the data is awkward. It handles inconsistent columns and half-clean exports sensibly, explains what a metric actually measures, and is unusually willing to say when a result looks unreliable. If the hard part of your problem is the thinking rather than the computing, that matters more than raw throughput.

Its ceiling is the tightest of the three: individual uploads are capped well below what a dedicated data tool accepts, and a conversation only holds so many files. Wide exports meet that wall sooner than people expect, and dropping unused columns before uploading helps more than compressing the file does.

Where does Gemini fit?

Gemini's strongest card is proximity. If your company runs on Google Workspace, the export is already in Drive, your team already lives in Sheets, and Gemini is right there with nothing new to adopt.

Two constraints shape it. Spreadsheet handling generally sits on the paid tier rather than the free one, so the first ceiling is often a subscription rather than a technical limit. And if you go the Sheets route rather than uploading, a Google Sheet only holds so much before it stops accepting rows — a large export from a payments or CRM system reaches that line faster than expected.

Is file size where the difference lies?

No. Paperswift's own file ceiling sits in the same range as these assistants', and Gemini's per-file allowance is larger than ours. If your only problem is that a file is too big to upload, swapping tools may not solve it.

The difference shows in rows rather than megabytes. Because the work happens on your own machine rather than in a shared sandbox with a time budget, a narrow file with a great many rows behaves very differently. That is a narrower claim than "we handle bigger files", and it is the accurate one.

What does none of them change?

The first is the upload. Whichever assistant you choose, your rows are sent to a vendor to be read. Each publishes clear commitments about how that data is handled, and for a great deal of work that is a reasonable trade. It stops being reasonable when the file is a customer list, a payroll run, or anything covered by an agreement about where data may travel.

Paperswift's answer is structural rather than contractual. The export is opened inside your own browser, and the only thing used when you ask a question is your column names and whether each holds text, numbers, or dates — never the values in any row. There is no upload step to trust because there is no upload step. The privacy page sets out where the line falls.

The second is repeatability. Ask any of the three about January's export and then February's, and the model may reason its way to the answer along a different path — usually the same destination, occasionally a slightly different definition. Paperswift users can save the query behind an answer and run it against an updated data file so you reuse verified and trusted logic again and again.

What does each one cost to set up and run?

None of them needs infrastructure, so the bill is a subscription rather than a server. The part that catches people out is not the subscription price. It is how fast spreadsheet work spends it.

All three meter usage, and a spreadsheet is an expensive thing to hand a model. The file has to be read into its context before a single question can be answered, and it is carried again with every follow-up, so a handful of questions about a wide export can eat more of your allowance than a fortnight of ordinary writing. If you already pay for one of these tools anyway, a quick look at a small CSV really may cost you nothing extra — but that is the best case, not the common one.

The trouble is that none of this is visible while you are doing it. Nothing on screen says the question you are about to ask is an expensive one; the feedback arrives afterwards, as a cap and a wait, usually partway through the task you were trying to finish. Someone fluent in context windows can work around it by dropping unused columns before uploading and starting fresh conversations. Someone who just wants last quarter's numbers has no way of knowing that attaching the larger file was the decision that cost them the rest of the afternoon — and that person is exactly who these tools are being sold to.

Free tiers compress the same problem into a smaller space: ChatGPT limits how many files a free account can upload each day, Claude caps free usage, and Gemini generally keeps spreadsheet uploads on the paid tier altogether.

Paperswift is metered differently, and mostly not at all. Questions are unlimited on every tier, including the free one, and that is architectural rather than generous — the model only ever receives your column names and their types, so a small file and an enormous one send the same short description. The work of running the query happens on your own machine, and re-running one you have already worked out involves no model call whatsoever. There is no way to spend an allowance by accident, because file size and model usage are not connected in the first place.

The free tier answers as many questions as you like and shows every chart on screen. Taking results back out (downloading, saving a chart, exporting) is what the paid tier adds. Current numbers live on OpenAI's, Anthropic's, and Google's pricing pages and on ours.

Which one should you use?

Use ChatGPT when you want a capable generalist and the data is not sensitive. Use Claude when the export is messy and the reasoning is the hard part. Use Gemini when your work already lives in Google Workspace and the convenience of staying there outweighs the rest.

Use Paperswift when the constraint is real: the file should not leave your machine, or the same report has to rebuild itself identically next month. It is much narrower than any assistant here — it will not draft your email or think through your strategy. For turning an export into a chart you can check, privately, that narrowness is the point.

Plenty of people sensibly use both kinds of tool, asking an assistant what to measure and Paperswift to measure it.