Can ChatGPT Do Math? I Checked Its Answers to the Cent
Short answer: for everyday sums, mostly yes — but only because the tools now quietly hand the arithmetic to code. I gave ChatGPT, Gemini and Claude a month of bakery expenses and every total came back right to the cent. Then I told ChatGPT not to use any tools and asked a compound-interest question. It answered to the cent, with no hedge, and it was $6.60 out.
That second result is the one worth knowing about, because the wrong answer looked exactly like the right ones.
How I set the test up
I wrote twelve invented expense lines for a small bakery — flour, butter, rent, utilities, card fees — the kind of list a small-business owner might actually paste into a chat. Before asking any of the tools, I worked out every correct answer myself in code, using exact decimal arithmetic. So each result below is checked against a known answer, not eyeballed.
Then I ran four prompts, each in a fresh chat:
| Test | Tool | Correct answer | What came back |
|---|---|---|---|
| Total, a percentage, and a what-if | ChatGPT and Gemini | $3,463.66 · 24.27% · $3,519.90 | All three right, both tools |
| Confirm my (wrong) total | ChatGPT | $3,436.66 is $27.00 low | Caught it |
| Messy figures from a bank app | Claude | $3,421.56 | Right, with an unstated assumption |
| Compound interest, no tools allowed | ChatGPT | $17,192.83 | $17,199.43 — wrong |
Test one: three questions, final numbers only
I asked for the monthly total, the share that went on ingredients to two decimal places, and the new total if only the flour bill rose 18%. I asked for the final numbers with no working, which is how most people ask.
Gemini and ChatGPT both returned $3,463.66, 24.27% and $3,519.90 — all correct. The interesting part is the small “Show code” link above Gemini’s answer. Gemini wrote a few lines of code, ran them, and reported what came out, rather than producing the total the way it produces a sentence. ChatGPT gave no visible sign either way of how it got there, which matters later.
Test two: will it just agree with me?
A more realistic risk than a botched sum is a chatbot that agrees with whatever number you hand it. So in a new chat I pasted the same expenses and said my bookkeeper had got $3,436.66 — the right digits with two of them swapped, the most ordinary typo there is — and asked ChatGPT to confirm it.
It said no, gave the correct total, and named the gap as $27.00, which is right. I half expected a polite yes. Even so, I would not lean on this: asking a tool to confirm your number invites it to agree, and one good run is not a habit you can rely on. Asking it to calculate and then comparing the two yourself is the safer shape of the same question.
Test three: figures copied out of a bank app
Real numbers rarely arrive clean. For this one I mangled the list the way an export does: dollar signs on some lines and not others, a comma in the rent, “287.9” with a missing zero, and a butter refund written in accounting brackets as (42.10).
Claude returned $3,421.56, which is correct. It read the brackets as a minus sign and subtracted the refund. That is the right call — and it made it without saying so. If those brackets had meant something else in your export, nothing in the answer would tell you a line had been taken away rather than added. I did ask for the number only, so that is partly my doing. It is also what most people type.
Test four: take the calculator away
Finally, I asked ChatGPT to work something out without running any code or using a calculator tool: a $12,847.39 balance growing 4.25% a year, compounded yearly, for seven years, to the cent.
It said $17,199.43. The correct figure is $17,192.83, so it was $6.60 high. On a five-figure balance that is small enough to look right, and stating it to the cent makes it sound checked. There is no “roughly”, no “about”, nothing to suggest it was an estimate.
This is what happens when a language model does arithmetic in its head. It is predicting the digits of an answer the way it predicts the words of a sentence, and long multiplication chains are exactly where that drifts. With a tool to run the sum, it is reliable. Without one, it produces something answer-shaped.
The catch: you do not decide when it uses code
In tests one to three, the tools chose to calculate properly. In test four, I stopped ChatGPT from doing so on purpose. In normal use, nobody tells it either way — the tool decides, and the decision can change with the model behind it, the plan you are on, or how the question happens to be worded. I covered how much of that moves under your feet in why ChatGPT gives different answers to the same question.
So the question to ask of any number it gives you is narrower than “can it do math?” — it is “did it actually calculate this one?”
How to get numbers you can rely on
- Ask it to calculate with code. Say so in the prompt: “Use code to work this out and show me the code.” When the tool shows you what it ran, you can see which numbers went in.
- Ask for the working, not just the result. A line-by-line breakdown lets you spot-check two lines in ten seconds. A bare number gives you nothing to check against.
- Make it state its assumptions. “List any line you treated as negative, skipped, or rounded.” That one sentence would have surfaced the refund in test three.
- Do not ask it to confirm your figure. Ask it to calculate independently, then compare the two numbers yourself.
- Re-add anything with money attached. A spreadsheet SUM takes seconds and does not guess. If the data already lives in a sheet, the approach in analyzing a spreadsheet with AI keeps the arithmetic where it can be checked.
Where I would trust it, and where I would not
For totals, percentages, splitting a bill, a quick what-if on a budget, or converting units, these tools are now good — especially when you can see that code ran, or when you are going to glance at the working anyway.
For interest, loan payments, tax, payroll, or anything where a few dollars of drift turns into a real cost, treat the answer as a draft. Check it the way you would check any confident claim — my two-minute fact-check routine works for numbers too.
What this test does not show
Four prompts, one run each, one invented budget. That is a demonstration, not an accuracy rate, and I would not claim any of these tools gets a particular percentage of sums right. I also did not test long word problems, statistics, or anything involving dates, where the ways to go wrong are different. The narrow finding holds: when the tools calculated, they were right every time, and the one time I stopped one calculating, it was wrong in a way that gave no warning.
Related reading
- How to Fact-Check an AI Answer in Under Two Minutes — the same habit, applied to anything stated with confidence.
- How to Analyze a Spreadsheet and Find Insights With AI — for when the numbers are already in a sheet.
- Why Does ChatGPT Give Different Answers to the Same Question? — why a tool that got it right yesterday might not today.
Written by Mitch, a software analyst who tests software for a living. This guide comes from actually using the tool on a real task. Tested on ChatGPT Plus, Claude Max and Gemini (free). More about this site · Corrections: acheatsheet@gmail.com.
Get the next guide by email
A new tested how-to when there is one — usually a couple a month. No hype, unsubscribe any time.
Tested in real accounts. No affiliate links.
