Claude vs ChatGPT for data analysis: which to trust with your numbers

Upload the same CSV to both and ask for last quarter's churn. You'll get two confident numbers, often different, and both will have been calculated by code the model wrote and ran. The question isn't which can do data analysis. Both can. It's which one makes fewer silent choices you'd disagree with, and which one tells you about them.

Sources: Claude help centre (uploading files; create and edit files) and OpenAI help centre (file uploads FAQ), checked 22 October 2026.

For analysis we'd pick Claude. Its models have led the independent finance-analyst benchmark, it handles long documentation alongside data, and it hands back finished files rather than a chat transcript, which matters more than it sounds: an analysis you can open and re-run is one you can check. We've argued that case in why AI analysis tools need to produce artifacts.

ChatGPT's best counter is practical. It accepts larger spreadsheets into its code environment, its interactive charts make exploring a dataset quicker, and it has run code on uploads for longer. If your work is poking at a 40MB export to see what's in it, that's a real advantage.

Two kinds of number circulate about AI and data. The flattering kind comes from narrow tests that are nearly solved. The useful kind comes from tests built on real, messy work, and those scores are much lower.

Sources: SpreadsheetBench official page (checked 22 October 2026); Adyen and Hugging Face, DABstep (arXiv, June 2025); Vals AI (updated 4 June 2026).

Read down the right-hand column. On the most realistic spreadsheet test, the best system finishes about a third of the workflows. On the finance-analyst test, the leader gets roughly one task in three wrong, and the leaders are Claude models from the generation before the current one. Vals now runs a second version of that benchmark whose scores aren't comparable, so check which one you're reading.

SQL tells the same story. On Spider 2.0, built from real enterprise databases, an o1-preview agent solved 21.3% of tasks against 91.2% on the original academic Spider (Lei et al., ICLR 2025). Specialised agent systems have since climbed the newer leaderboards, but the gap between textbook and production schemas hasn't gone away.

None of these benchmarks tests the Claude or ChatGPT apps directly. They test models and agent systems. Treat them as a ceiling on what to expect, and see the most accurate AI for data analysis for the wider field.

No benchmark can tell you whether this answer, on your data, is right. Count can: connected to Claude, it shows the query or calculation behind each number and flags the ones that rest on an assumption.

Three customers pay $20, $30 and $50 a month. The $50 customer leaves. What's your churn rate? By customer count it's 33%. By revenue it's 50% (ChartMogul's worked example). Both are correct. A model that gives you one of them without saying which has made a business decision on your behalf.

That's the failure to worry about most, and it isn't arithmetic. Code execution has largely fixed the arithmetic: older models doing sums in their heads were poor at it, with GPT-4 managing 59% on three-digit multiplication (Dziri et al., 2023), which is why both vendors now route calculations to code. What code can't fix is the choice of formula.

• Churn rate: customers or revenue; gross or net (net can go negative); Stripe's denominator includes subscribers who joined during the period, ChartMogul's uses customers at the start.
• ROAS: Google Ads defines it as conversion value divided by spend, shown as a percentage, so 5x and 500% are the same result. Conversion value isn't always cash revenue, and refunds may or may not be netted off.
• Monthly active users: GA4 counts a rolling 30 days, not a calendar month, and its "Users" means active users.
• Customer lifetime value: revenue or margin, discounted or not, 120 days (GA4) or several years.

One more, and it's the sneakiest. Ask for ROAS across a month of daily rows and a model may average the daily ratios. Illustrative numbers: day one, $100 spend and $500 revenue (5x); day two, $1,000 spend and $1,000 revenue (1x). The average of the ratios is 3x. Total revenue over total spend is $1,500 / $1,100, about 1.36x. Same data, very different story for the board.

Irrelevant detail makes it worse. Apple researchers found that adding one plausible but irrelevant clause to a maths problem cut accuracy by up to 65% (GSM-Symbolic, 2024). Real exports are full of irrelevant columns.

Build a small table you know the answer to. The point is to set traps a careful analyst would spot. Everything below is illustrative; use your own numbers if you have them.

The data: ten subscription customers, with monthly price at 1 March and at 31 March. Include one who cancelled, one who downgraded from $50 to $20, one who joined and cancelled within March, and one who returned after cancelling in February. Work out logo churn and revenue churn yourself first.

The prompt, identical in both:

Here is our subscription data for March. What was our churn rate?
Show the code you ran.

What to score, for each model:

1. Did it run code, or answer from reading the table?
2. Did it say which churn it calculated (customers or revenue) without being asked?
3. How did it treat the downgrade? Contraction, churn, or ignored?
4. What happened to the customer who joined and left in March? In the denominator, the numerator, both, neither?
5. Did it mention the returning customer, and whether it netted them off?
6. Run it again in a fresh chat. Same number?

Then ask: "What assumptions did you make that aren't stated in the data?" A strong answer lists the four traps. A weak one restates the result.

We're not going to tell you which model wins, because it varies with phrasing and from run to run, and you'll learn more from your own result. What the test shows is whether a model surfaces its choices. For prompts that make it more likely to, see prompts for data analysis.

An analysis has more places to hide an error than a paragraph does: the rows it loaded, the filter it applied, the join, the formula, the rounding. You can't check all of that every time. You can check the parts that fail most often.

A two-minute check

1. Row count. Ask how many rows it read. Compare with your file. Dropped header rows and silently skipped blanks show up here.
2. Definitions. "Define every metric in your answer in one line each." Compare with the definition your team uses.
3. One number by hand. Pick the headline figure and recompute it in a spreadsheet from the totals.
4. The code. Skim it for a filter you didn't ask for or a date range that's off by one.
5. A second run. If the headline number changes, the analysis isn't settled.

Doing that for every chart is where analysts' time goes. Count was built for this: connect it to Claude or ChatGPT and it splits the answer into claims, each with a reasoning chain showing the query or calculation underneath and the assumption it rests on. Its canvas also runs SQL and Python against your warehouse directly, so the analysis and its checks live in one place.

For more on each tool's data features, see Claude data analysis.

Is Claude or ChatGPT better for data analysis?

What is the best AI for data analysis?

Can ChatGPT analyse Excel files?

Can Claude analyse data?

How accurate is AI at data analysis?