Which AI is best at Google Sheets?
The same Google Sheets tasks, given word for word to every model. Run by spreadsheet people in their own AI, graded in their own sheets.
Sample leaderboard: these numbers are made up to show how it will look. Real scores replace them as soon as the first results are checked. Add the first result
Sheets score (sample)
% of formula and Apps Script tasks passed (the two averaged) · 17 tasks
- Anthropic
- OpenAI
- DeepSeek
- xAI
- 1.Claude Opus 5.5 98
- 2.GPT-5 94
- 3.Claude Sonnet 5.5 92
- 4.Claude Haiku 5.5 88
- 5.DeepSeek V3 88
- 6.Gemini 3.1 Pro 86
- 7.Grok 4 73
- 8.GPT-5 mini 71
- 9.Gemini 3.8 Flash 58
Hover a bar for its category scores. Full table below.
Score vs cost (sample)
Sheets score against the average API cost of one task. Up and to the left is better and cheaper.
Scores by category (sample)
Formulas and Apps Script are answers you paste into a sheet. Computer use and browser extensions are AI agents doing the work in Sheets themselves.
| Model | Formulas | Apps Script | Computer use | Browser extensions | People's pick | Overall |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 96% | 100% | 100% | 100% | 50% | 98 |
| GPT-5 | 88% | 100% | 33% | 100% | 47% | 94 |
| Claude Sonnet 5.5 | 83% | 100% | 0% | 67% | 56% | 92 |
| Claude Haiku 5.5 | 75% | 100% | 0% | 67% | 52% | 88 |
| DeepSeek V3 | 75% | 100% | not tested | not tested | 45% | 88 |
| Gemini 3.1 Pro | 71% | 100% | 100% | 67% | 48% | 86 |
| Grok 4 | 79% | 67% | not tested | not tested | 46% | 73 |
| GPT-5 mini | 75% | 67% | 33% | 33% | 52% | 71 |
| Gemini 3.8 Flash | 83% | 33% | 33% | 67% | 55% | 58 |
Run a test in your own AI
Pick a task, copy the prompt exactly as written, paste it into ChatGPT, Claude, Gemini or whatever you use. Then paste the answer back here and say whether it worked in your sheet. Every answer is checked before it counts.
Agent tasks
These are run by Better Sheets in each AI's computer-use agent or browser extension, starting from the same sheet every time, and checked by reading the finished sheet.
- Computer use Build an invoice from a blank sheet Turn this sheet into a reusable invoice with line items, quantity, price, a tax rate cell, subtotal, tax and total, formatted as currency.
- Computer use Clean a messy CSV import Clean this data so every date is a real date, text has no extra spaces, and there is only one header row.
- Computer use Make a monthly chart Add a column chart of total sales per month on a new tab called Chart.
- Browser extensions Highlight overdue rows Highlight every row where Due is before today and Status isn't Done, and keep it updating automatically.
- Browser extensions Add status dropdowns Make the Status column a dropdown with To do, Doing, Done, and colour each option.
- Browser extensions Pull a table from a web page Put the table of the 10 largest countries by area from this page into the sheet, with numbers as numbers.
How the benchmark works
- Every model gets the exact same prompt
- Tasks are frozen and numbered. If a prompt ever changes, it becomes a new version and old scores aren't mixed with new ones.
- Pass or fail, checked in a real sheet
- An answer passes when it gives the right result in Google Sheets, not when it looks right. Every pasted result is reviewed before it counts.
- Some tasks are kept secret
- A few tasks are never published, so a model can't do well just by having seen the answers.
- "People's pick" is a blind vote
- Two right answers can still differ: one is shorter, easier to edit, or survives new rows. Voters pick without knowing which AI wrote which.