You want to know which AI to trust with your spreadsheets. ChatGPT, Claude and Gemini all have a sidebar inside Google Sheets now, Excel has Copilot, and a dozen startups say theirs is best. Search “best AI for spreadsheets” and page 1 is mostly tool lists. None of them show you a single score.
The scores exist. In the last twelve months researchers, Microsoft, independent testing firms and a handful of startups have published at least a dozen spreadsheet benchmarks. I read all of them so you don’t have to. Here’s what they found, where they disagree, and the one thing none of them test: Google Sheets formulas and Apps Script.
TL;DR
- The best AI now scores 83.11% on SpreadsheetBench, but on realistic multi-sheet workbooks the best bare model solves only 34.89%.
- Every benchmark I found tests Excel files, so nobody is publicly scoring Google Sheets formulas or Apps Script yet.
- The most common mistake is trusting the AI when it says it’s done: 80.7% of Claude’s WTMBench failures were exactly that.
Who this is for
- Google Sheets power user: You’ll see why a big SpreadsheetBench number says nothing yet about QUERY, ARRAYFORMULA or Apps Script.
- Financial analyst: The best AI-built finance sheets pass 89% of formula checks but only 64% of number checks, so you know what to recheck.
- Operations manager: You’ll learn which jobs AI handles well today and why pivot tables and conditional formatting still need a human.
- Excel consultant: You get the real scores behind the launch posts, so you can tell clients what Copilot, Claude and ChatGPT add-ins can and can’t do.
- Team lead choosing an AI tool: You’ll know why a one or two point gap between tools is noise and why the wrapper matters as much as the model.
The short version: every benchmark, best score
- SpreadsheetBench (912 tasks): Real questions from Excel forums. Best: 83.11%, Kingsoft’s Qingqiu Agent (Jun 2026).
- SpreadsheetBench Verified (400 tasks): A cleaned-up subset of the same. Best: 99.25%, two agents tied (Sep 2026).
- SpreadsheetBench 2 (321 tasks): Multi-sheet business workbooks. Best: 54.51%, arito; best bare model 34.89% (Sep 2026).
- WTMBench (Microsoft): Formulas, charts, pivot tables, conditional formatting. Pivot tables succeed only 0 to 10% of the time (Aug 2026).
- Vals AI Excel Modeling: LBO, DCF and operating models. Best: 76.67%, Claude Fable 5.1 (Oct 2026).
- AA-AnalystAgent: Analysis over folders of spreadsheets. Best: 54% on all 5 tries, Claude Opus 5 (Aug 2026).
- Finch: Real finance and accounting workflows. Best: 38.4%, GPT-5.1 Pro (Dec 2025).
- SpreadsheetArena: People vote on AI-built workbooks. Claude Opus 4.5 was on top in the paper’s January 2026 snapshot.
- BlueFin: 131 professional finance tasks. The best models average under 50% (May 2026).
Read that list top to bottom and you see the whole story. On the famous benchmark, AI looks superhuman. The closer a test gets to a real workbook someone would actually send to their boss, the lower the scores go.
SpreadsheetBench: the one everyone quotes
When Google, Microsoft or a startup puts a spreadsheet number in a launch post, it’s almost always SpreadsheetBench. It came out of Renmin University in 2024: 912 real questions people asked on Excel forums, each with a spreadsheet, and a checker that compares the AI’s output cells to the right answer on about three versions of the sheet with different numbers. It’s open source, so anyone can run it.
Three things about it surprised me.
The “human expert” score that AI beat came from 4 people. The paper’s human baseline is 71.33%. That number is four Excel experts working through 50 of the tasks in one day (paper, Table 2). When it was published, the best AI, GPT-4o, scored 18.35%. Now Kingsoft’s Qingqiu Agent scores 83.11% on the full set, China Mobile’s JT AlphaData 77.85%, and Google’s own Gemini in Sheets 70.48%. Microsoft’s Agent Mode in Excel launched at 57.2%. The top of the board isn’t an American AI lab. It’s Chinese office suites.
More than half of it had to be fixed. When Shortcut and the benchmark authors built SpreadsheetBench Verified in December 2025, they removed 216 of the 912 tasks and modified 382 more. The reasons will sound familiar to anyone who answers spreadsheet questions: the question was unclear, there was more than one right answer, or the answer used RAND or TODAY and changed every time.
The cleaned-up version is already maxed out. Two agents hit 99.25% on Verified in September 2026. And the noise is as big as the gaps: GRID re-ran Claude models on it and found 12 to 14 tasks out of every 100 flipped between pass and fail from one run to the next, even with randomness turned all the way down. When two tools are one or two points apart, that’s a coin flip.
Is it in the training data? The questions are public forum posts, and the full dataset has been open since April 2025. The authors rewrote questions and changed the data to make memorizing harder, but I couldn’t find anyone who has checked whether today’s models have seen it. Keep that in mind when a launch post leads with it.
The harder tests tell a different story
Real workbooks: about 1 in 3. SpreadsheetBench 2 came out in June 2026 with 321 tasks built like real work: filling templates, financial models, debugging someone else’s sheet, charts. The best plain model, Claude Opus 4.6, solved 34.89%. Debugging tasks went as low as 12%. The authors also tested four of the add-ins you can install today on 30 of the tasks: Kimi Sheet, GLM in Excel, Claude for Excel and ChatGPT for Excel. The best one, Claude for Excel, scored 15.4%, and none of the products beat the raw models (paper). The most common mistakes had names I loved: “Insufficient Inspection” and “Wrong Target Selection.” The AI didn’t look at the sheet before changing it, or changed the wrong part.
Pivot tables are still almost unsolved. Microsoft’s own research group (the people behind Flash Fill) built WTMBench to test the things people actually click around to make. Pivot tables: 0 to 10% success. Conditional formatting: 2 to 36%. Charts about 50 to 61%, formulas about 32 to 65%.
The formulas look right more often than the numbers are right. Vals AI’s Excel Modeling Benchmark has AI agents build LBOs, DCFs and operating models. The top model, Claude Fable 5.1, passes 89% of the checks on formulas but only 64% of the checks on the actual numbers (updated October 16, 2026). The structure is there. The answer is off.
Only 1 in 6 AI-built finance sheets was good enough to send. In SpreadsheetArena, five finance experts rated 50 AI-built workbooks. They averaged 2.86 out of 5, and only 16.1% were rated good enough to send to a client or colleague. Formatting and color coding was the weakest part. The experts agreed with the crowd’s votes only 56% of the time, which says a lot about “people liked it” as a measure.
Five ways AI gets spreadsheets wrong (with numbers)
- It says it’s done when it isn’t. In WTMBench, 80.7% of Claude’s failures and 43.2% of GPT’s were what the researchers call hallucinated successes: the model reported the task finished, and it wasn’t.
- It types the answer instead of a formula. In MBABench, GPT-5.4 hard-coded 24.7% of the answer cells. The sheet looks right today and is wrong the moment an input changes. Rows found the same thing in its own test: only 6% of Gemini in Sheets’ answers updated when the data changed.
- It gives a different answer each time. On Artificial Analysis’ test, Gemini 3.1 Pro solved 81% of tasks at least once but only 41% on all five tries.
- It doesn’t read the sheet first. That’s the “Insufficient Inspection” failure that topped SpreadsheetBench 2.
- It’s only as good as the wrapper around it. The same base model went from 80.25% to 91.25% on SpreadsheetBench Verified once Leni wrapped its own tools, checks and routing around it (Leni paper). When a launch post compares “our product” to “Claude” or “GPT”, it’s often comparing a whole system to a bare model.
“If there’s a shortcut to be taken, an AI agent will take it.” (Ian Schnoor, Financial Modeling Institute, ICAEW)
He was describing an agent that wrote 1,000 formulas in a second, full of hard-coded numbers. That matches the research above.
What about AI formulas inside the cell?
Both big spreadsheet apps tried putting AI straight into a formula.
Google’s =AI() function launched in June 2025. It’s useful for text: summarizing, sorting into categories, pulling sentiment out of reviews. Per Google’s help page it fills at most 350 cells per run (it launched at 200), can’t see the rest of your spreadsheet, can’t be undone, and can’t be nested inside IF.
Microsoft’s =COPILOT() function arrived in Excel’s beta in August 2025. Microsoft’s own support page told people not to use it for numerical calculations. It never left preview, and Microsoft removed it on September 14, 2026.
My take: an AI formula is fine for messy text. For numbers, have the AI write a real formula you can read.
Common pitfalls
- Trusting the score in a launch post: Check which benchmark it is. The cleaned-up SpreadsheetBench is maxed out at 99.25%, and realistic workbooks still sit around 1 in 3.
- Accepting hard-coded values: Click a few answer cells and make sure there’s a formula, not a typed number. Then change an input and see if the answer moves.
- Not checking pivot tables and conditional formatting: These scored 0 to 10% and 2 to 36% in WTMBench, so build or at least review them yourself.
- Believing “done”: Open the sheet and look before you move on. Most of Claude’s failures in WTMBench were tasks it said it had finished.
- Judging an AI on one run: Run the same task a few times. Gemini 3.1 Pro solved 81% of tasks at least once but only 41% on all five tries.
- Using =AI() for numbers: Keep in-cell AI for messy text. For numbers, ask the AI to write a real formula you can read and check.
What none of these benchmarks test
Every academic benchmark above is Excel. The tasks are .xlsx files, and SpreadsheetBench’s checker recalculates them in Excel or LibreOffice. Gemini in Sheets’ 70.48% was measured on those same Excel tasks.
So nobody is publicly testing:
- Google Sheets formulas. QUERY, ARRAYFORMULA, FILTER and SORT that spill, IMPORTRANGE, REGEXEXTRACT, LAMBDA. The formulas Sheets people actually use.
- Apps Script. Writing it, fixing it, triggers, custom menus, quotas. Claude for Google Workspace, which launched October 6, 2026, can’t create triggers or custom menus at all (Anthropic’s help page).
- The three sidebars head to head. Gemini, ChatGPT for Sheets and Claude now all run inside Google Sheets. The only comparisons are vendors testing their own product.
- How often the same AI gets the same answer twice. GRID and Artificial Analysis showed that matters, and almost nobody reports it.
That’s why I started the Better Sheets AI Benchmark. Every model gets the same frozen Google Sheets prompts, word for word: formulas, Apps Script, and tasks for computer use agents and browser extensions. You can run a task in whatever AI you already use and paste the answer back, and grade other people’s answers without knowing which AI wrote them. A few tasks are kept secret so no model can learn the answers. Right now the leaderboard shows sample numbers so you can see how it’ll look. Real scores replace them as soon as the first results are checked.
Learn the Google Sheets formulas in this post
- 3 Ways to Count in Google Sheets: count rows yourself so you can check the totals an AI gives you.
- How to Find Duplicates: catch the repeated rows an AI cleanup can quietly leave behind.
- How To Sort A Filter: understand the FILTER and SORT formulas AI likes to write in Google Sheets, so you can read them.
- Most Common Errors and How To Handle Them in Google Sheets: fix the #N/A and #REF errors that show up when an AI formula points at the wrong place.
- 200 REGEXMATCH Examples in Google Sheets: check and tweak the REGEX formulas an AI writes for messy text.
Better Sheets has hundreds more tutorials like these. If you want to get better at Google Sheets, join Better Sheets.
FAQ
What is the best AI for spreadsheets right now?
It depends on the test. On the full SpreadsheetBench, Kingsoft’s Qingqiu Agent leads at 83.11%, and on the harder SpreadsheetBench 2 the best product, arito, scores 54.51%. Nobody has a public answer for Google Sheets yet.
Is SpreadsheetBench reliable?
It’s useful but limited. More than half of the original tasks had to be fixed or removed, the cleaned-up version is maxed out at 99.25%, and 12 to 14 tasks out of every 100 flip between pass and fail from one run to the next.
Can ChatGPT, Claude or Gemini write Google Sheets formulas?
All three now run in a sidebar inside Google Sheets. But no public benchmark tests Google Sheets formulas, so there’s no independent score yet. That’s what the Better Sheets AI Benchmark is for.
What happened to the COPILOT function in Excel?
Microsoft’s =COPILOT() arrived in Excel’s beta in August 2025, never left preview, and was removed on September 14, 2026. Microsoft’s own support page had told people not to use it for numerical calculations.
How many cells can =AI() fill in Google Sheets?
At most 350 cells per run, up from 200 at launch. It can’t see the rest of your spreadsheet and can’t be undone.
Is there a Google Sheets AI benchmark?
I couldn’t find a public one for Google Sheets formulas or Apps Script. Gemini in Sheets’ 70.48% was measured on SpreadsheetBench’s Excel tasks. The Better Sheets AI Benchmark is my attempt to fill that gap, and right now it shows sample numbers until the first real results are checked.
