Is ChatGPT Accurate with Numbers? Where It Is and Isn't

Is ChatGPT accurate with numbers and calculations? When it runs code, it is as accurate as the calculator on your desk. When it answers from its head, it is usually right on short arithmetic and unreliable on long or multi-step work. And in our experience the arithmetic is rarely the real problem. Far more wrong numbers come from a wrong input or a wrong formula, calculated perfectly. This article shows how to tell the two modes apart, the four ways a number goes wrong, and a prompt that makes the calculation checkable.

Why a language model can miss simple math

A language model produces text one piece at a time, choosing each piece by what is likely to come next. A long number is split into chunks, and there is no built-in register carrying the 1 from one column to the next. For 12 × 12 the likely continuation and the correct one are the same thing. For 48,317 × 6,209 there is no memorized answer, and the model has to simulate the procedure in text, where one slip in the middle changes the result while the output still looks like a number of the right size.

Reasoning models, which work through steps before answering, have made in-head arithmetic much better than it was two years ago. Better is not the same as guaranteed, and the answer arrives in the same confident tone either way.

Framing matters too. In 2024 a group of Apple researchers published GSM-Symbolic, a test built from school-level word problems. Changing only the numbers in a problem changed how often models solved it, and adding one irrelevant sentence cut performance by up to 65 percent in the models tested at the time. Today's models are stronger. The practical lesson stands: how a problem is worded can move the answer, which never happens with a spreadsheet.

Two modes: calculating and recalling

ChatGPT and Claude can write a few lines of Python and run them, and Google documents the same code-execution ability for its Gemini models. The number you get then comes from the computer, and the model only reports it.

How to tell which mode you got: look for the code. ChatGPT shows an analysis step you can expand. Claude shows the script it ran when code execution is switched on (Settings → Capabilities; available on every plan as of October 2026). If you see a result and no code, the model did it in its head.

You do not have to hope for the right mode. Ask for it: "Use Python for this calculation and show the code." It is the single most useful sentence for numeric work.

The four ways a number goes wrong

1. Arithmetic slips. Long multiplication, many chained steps, large numbers done without code. This is the error everyone worries about, and since the tools started running code it has become the least common one in our own work.

2. Wrong inputs. The model misreads a figure from a PDF or a screenshot, swaps two digits, mixes thousands with millions, or quietly uses a number from memory because you did not supply one. An exchange rate "remembered" from training data is a typical case. The calculation is flawless and the result is wrong.

3. Wrong method. The formula does not match what you meant. Three examples that come up constantly:

• Margin and markup. A product costs $80 and sells for $100. Markup is 25 percent (20 ÷ 80). Margin is 20 percent (20 ÷ 100). Ask for "the margin" loosely and you may get either.
• Percent and percentage points. Conversion rising from 2 percent to 3 percent is an increase of one percentage point and of 50 percent. Both are correct and they tell very different stories on a slide.
• Average of averages. Team A closed 10 deals averaging $1,000, team B closed 90 averaging $2,000. Averaging the two averages gives $1,500. The true average deal is $1,900, because (10 × 1,000 + 90 × 2,000) ÷ 100 = 1,900.

Code does not protect you here. Python will compute the wrong formula to as many decimal places as you like.

4. Invented numbers. A statistic, a benchmark, or a market size that the model produced because the sentence needed one. This is a sourcing problem and has its own routine: How to Fact-Check AI Answers.

Knowing which of the four you are guarding against tells you which check to run.

Trust, verify, or do it yourself

Situation	Verdict	Why
Sums, averages, and counts on an uploaded file, with code shown	Trust the arithmetic, check the filters	The computer did the math; the definitions are yours to confirm
Percentages, ratios, and unit conversions from inputs you typed, with code shown	Trust	Easy to reverse-check in seconds
Any arithmetic with no code shown	Verify	It may be right; you cannot know
Numbers pulled out of a PDF, a scan, or a screenshot	Verify every figure you will use	Misreads look exactly like correct reads
Date math: days between, deadlines, business days	Verify	Off-by-one and calendar rules are common failures
Currency conversion, interest rates, prices	Verify the rate, not the multiplication	The input goes out of date
A total the model carries through a long chat	Verify	Figures drift as the conversation grows; restate them
Tax, payroll, invoices, statutory rates	Do it in the proper system	Rules change and the mistake is yours
Statistics quoted from memory	Never use unverified	See error type 4
Medical doses, engineering loads, anything safety-related	Never rely on a chatbot alone	The cost of one error is too high

For whole spreadsheets the same logic applies with a few extra steps, described in How to Analyze a Spreadsheet with ChatGPT.

A working example: campaign numbers for a report

A typical task. A marketing manager is closing the month and needs five figures for a report: cost per lead, customer acquisition cost, conversion rate, revenue, and return on ad spend. The inputs are in an email from the ads team.

The weak version of the request is "calculate CPL, CAC and ROAS from these numbers." It will usually work. When it does not, you have nothing to check it against. The stronger version fixes the mode, the inputs, and the format of the answer.

Use Python for every calculation in this chat. Do not compute anything in your head.

Inputs. Treat these as the only source of numbers; do not add figures from memory.
- Ad spend: $18,400
- Leads: 1,150
- New customers: 92
- Average first order: $640

Calculate:
1. Cost per lead
2. Customer acquisition cost
3. Lead-to-customer conversion rate
4. Revenue from first orders
5. Return on ad spend (revenue divided by spend)

For each result give: the formula in words, the inputs used, the result, the unit.
Then run one reverse check per result (for example, acquisition cost
multiplied by customers should equal spend) and say whether it passed.
If an input is missing or could be read two ways, stop and ask before calculating.

What a correct answer contains: cost per lead $16.00, acquisition cost $200.00, conversion 8.0 percent, revenue $58,880, return on ad spend 3.2. The reverse checks are simple enough to confirm in your head: 200 × 92 is 18,400, and 16 × 1,150 is 18,400.

Where it needs a follow-up:

• No code block appeared. Reply: "You did not run code. Run it and show it."
• Your finance team reports ROI, not ROAS. ROI here is (58,880 − 18,400) ÷ 18,400 = 2.2, or 220 percent. The inputs are identical and the headline number is different, which is error type 3 waiting to happen. Put the formula in the prompt, as we did in item 5.
• The model rounded early. If the parts are rounded before they are summed, the total can be off by a few units. Ask for full precision in the calculation and rounding only in the final table.

The same structure works for a loan schedule, a price increase, or a headcount budget: mode, inputs, formulas in words, reverse check. The general logic of prompts built this way is in How to Write Better ChatGPT Prompts for Work.

A five-step check for any number

• Source of each input: did I type it, did it come from a file I supplied, or did the model bring it?
• Mode: is the code visible?
• Size: before reading the answer, I estimate the order of magnitude. Spend of about 18,000 over about 1,000 leads should land near 18.
• Reverse: I undo the operation. Multiply back, add the parts, check that the shares sum to 100 percent.
• Independent recompute: the one number that will be quoted goes through a spreadsheet or a calculator.

The third step is the cheapest and catches the worst mistakes, the ones that are wrong by a factor of ten or a thousand.

Where this breaks

Code proves the arithmetic, not the method. If the script filters the wrong rows or divides by the wrong base, it runs without complaint. Read the formula in words. That is why the prompt asks for it.

Extraction from documents. A table in a scanned PDF is an image, and digits get misread. A 3 becomes an 8, a decimal point disappears. Check extracted figures against the page before any calculation.

Ambiguous business terms. Revenue can be booked or collected, a "customer" can be new or active, a month can be calendar or billing. The model picks one reading and does not mention it unless told to ask.

Long conversations. By the fortieth message a figure from the fifth may have changed by a digit. For work that spans a long chat, paste the confirmed inputs again before the final calculation.

Regulated calculations. Tax, payroll, and statutory interest depend on rules that change by date and by place. A chatbot can explain the logic. The filing belongs in the proper software or with an accountant.

Take this further in the practicum

Numbers get their own block in Session 3 of AI system in two weeks, the session on documents, data, and visuals: where AI is reliable with figures, where it is confidently wrong, and how to set up calculations so that every result can be traced. Participants work on their own reports and budgets, and the answer-verification checklist from the course materials covers numeric checks alongside sources and quotes. The full program and the upcoming cohorts are on the course page.