FOWL AI Tests · Monthly · Test #4

Four AI assistants built the same budget spreadsheet. All four got the math right.

ChatGPT, Claude, Gemini and Microsoft Copilot got the same freelancer budget brief, word for word, and each was asked for a real Excel file. We downloaded every file and recalculated every formula ourselves. The totals matched. What was underneath them did not.

Tools testedChatGPT (GPT-6.1 Sol), Claude (Opus 5.5), Gemini (free tier), Microsoft Copilot (free)

Test caseOne-page Q4 budget for a fictional freelance designer

What we checkedThe real .xlsx file: every formula, date and print setting

Key Findings: 15-Second Version

All four delivered a real Excel file, and all four ledgers end Q4 at the correct $5,667.50.
Claude built the sheet you can safely edit: the tax rate and starting balance are input cells, dates are real dates, and a built-in check confirms the summary matches the ledger.
ChatGPT was right behind it: formulas everywhere, a tax-rate input cell, set to print on one page.
Copilot typed the tax amounts in as plain numbers, so change an invoice and its tax row stays wrong. Gemini's "Ending Cash Balance" row shows errors instead of numbers.

The Setup

We invented a freelance graphic designer with a simple quarter: a monthly retainer, three project invoices, five monthly bills, one laptop purchase, and 25% of every payment set aside for taxes. Each assistant got the exact same prompt in a normal account, on its default setup. No follow-up prompts and no edits: we took the first file each tool produced.

We worked out the right answers before running the test: income of $14,450, $3,612.50 set aside for tax, November slightly negative (minus $140) because of the laptop, and an ending balance of $5,667.50. Then we opened every file, read each formula, and recalculated each workbook two ways: in Apple Numbers and with an open-source formula engine.

FreelancerFictional, invented for this test

PromptIdentical across all four, word for word

Output checkedThe downloaded .xlsx file, not the chat reply

VerificationEvery formula recalculated in two engines

Show the exact prompt we used

I'm a freelance graphic designer. Build me a one-page Q4 2026 budget spreadsheet (October, November, December) as a downloadable Excel file (.xlsx). Starting balance on Oct 1: $4,200. Income: a client retainer of $3,000 on the 1st of each month, plus project invoices paid Oct 15 ($1,850), Nov 20 ($2,400) and Dec 10 ($1,200). Monthly expenses: rent $1,450 on the 1st, software subscriptions $120 on the 5th, health insurance $380 on the 10th, phone and internet $140 on the 15th, groceries $500 on the 20th. One-off: a new laptop for $1,600 on Nov 28. Set aside 25% of every income payment for taxes on the day it arrives, as its own expense row. Include a transaction table with date, description, category, income, expense and a running balance column that uses formulas (not typed numbers), plus a monthly summary (total income, total expenses, tax set-aside, net) and Q4 totals, all formula-driven. Keep it to one printable page.

Fine print. Gemini ran on the free tier because our paid Google plan had lapsed, and Copilot ran on a free personal account; paid plans may behave differently. We recalculated the files in Apple Numbers, not Microsoft Excel. Copilot and Gemini store their dates as text, and Numbers handled that well; Excel can be stricter with text dates, so your results there may differ. We did not test that.

The Scoreboard

Five categories, scored from the actual files.

Every category is scored 1 to 5, based only on what we found in the real file, never the chat reply.

5 = Excellent 4 = Good 3 = Acceptable 2 = Weak 1 = Poor

Category	What it measures
Real Deliverable	Did it hand over a real .xlsx file that opens?
Math Accuracy	Every summary and balance cell, recalculated by us. The right ending balance is $5,667.50.
Formula-Driven	Running balance, tax rows and summaries built from formulas, with the 25% rate and starting balance as editable inputs.
Built to Use	Real dates (not text), set to print on one page, clear layout.
Honesty	Do its claims about its own file match the file?

Category	Claude	ChatGPT	Copilot	Gemini
Real Deliverable
Math Accuracy
Formula-Driven
Built to Use
Honesty
Total (out of 25)	25 Winner	24	16	16

Summary Showdown

What each tool said about its own file.

Verbatim from each chat reply, then what we actually found.

Claude

"I recalculated it, checked it against my own math, and it has no formula errors."

True. It also said the balance never drops below $4,500. We checked: the low point is exactly $4,500, on Oct 10.

ChatGPT

"It's formatted to fit one printed page, and the formulas were checked for errors."

True. Fit to one page is switched on, and it has all 28 transactions it said it had.

Copilot

"Automatic 25% tax set-aside entries for every income payment"

Not quite. The tax rows are typed-in numbers, not formulas. It also promised a layout "designed for printing on one page," but fit to one page is not switched on.

Gemini

"dynamic Excel formulas (SUMIFS, INDEX, MATCH) calculating ... ending cash balances"

The ending cash balance row is the one row that fails. It returns errors in both engines we tried.

Sample View

The actual spreadsheets. Open them yourself.

Claude Correct, editable

56 formulasReal datesRate + balance input cellsFits 1 page

Download the .xlsx →

What it did differently: Put the 25% rate and the $4,200 starting balance in two highlighted input cells, so changing either one updates every row. It added a "Matches ledger?" cell that checks its own summary against the transaction table, and a note explaining that the summary's expenses leave out tax.

ChatGPT Correct, editable

54 formulasReal datesTax-rate input cellFits 1 page

Download the .xlsx →

What it did differently: Added an ending balance for each month to the summary, and showed a tidy snapshot in the chat too. One small flaw: its note under the title is copied into six cells across the row instead of one.

Copilot Typed tax rows

45 formulasDates stored as textNo rate cell1-page print not on

Download the .xlsx →

What it did differently: Every number is right today. But the tax rows are fixed amounts, so if an invoice changes, its 25% set-aside does not. The summaries sit in three blocks beside the ledger, spread across 16 columns.

Gemini One broken row

57 formulasDates stored as text25% typed into each formulaFits 1 page (landscape)

Download the .xlsx →

What it did differently: The ledger and monthly totals are right, and it set up landscape printing on one page. Its "Ending Cash Balance" row looks up real dates in a column of text dates, so it finds nothing and shows an error in every month.

What We Found

Getting the math right is no longer the hard part

All four ledgers landed on $5,667.50 and all four caught that November goes negative. If your only question is "can AI add up my budget," the answer is yes.

The difference is what happens when you change a number

A budget is something you edit. In Claude's and ChatGPT's files, change an invoice or the tax rate and everything updates. In Copilot's, the tax rows are typed in, so they quietly go stale. In Gemini's, the 25% is written into each formula, so changing the rate means editing every tax row. Before you trust an AI spreadsheet, change one input and watch what moves.

Text that looks like a date is not a date

Copilot and Gemini both stored "2026-10-01" as text. It looks identical, but sorting, filtering and date lookups can treat it differently. That is exactly what broke Gemini's ending balance row. Claude and ChatGPT stored real dates.

Two tools described features their files did not have

Claude's and ChatGPT's descriptions of their own files checked out line by line. Copilot called its tax rows automatic and its layout one-page; neither was true in the file. Gemini listed ending balances among its working formulas; that is the row that fails. Same lesson as every test so far: read the file, not the reply.

🏆 The Verdict

Best overall

Claude

Correct, fully editable, real dates, and a built-in check that the summary matches the ledger. Every claim it made about the file was true.

Close second

ChatGPT

Just as correct and almost as editable, with a handy month-end balance in the summary. One point behind on a small layout flaw.

The gap between Claude and ChatGPT is one point. The real gap is between files you can keep editing and files that are only right until you change something. We tested one budget on default settings, so treat this as one honest data point, not the final word.

Should you switch tools?

Use Claude or ChatGPT if you will keep updating the sheet over the quarter.

Use Copilot for a quick one-time snapshot, then ask it to turn the tax rows into formulas.

Use Gemini if you want a landscape one-pager, then delete or fix the ending balance row.

We test a new AI tool matchup every month.

Real files, real numbers, no sponsorships. Subscribe to get the next one first.

Subscribe free →

All tests Weekly briefings Browse AI jobs