WTM-Bench (Python): leaderboard
WTM-Bench (2026) gives a model a public Excel workbook with its derived artifacts removed and a concise level-3 request, and asks it to recreate the formulas, charts, pivot tables or conditional formatting by writing code in a tool loop of up to 10 turns; outputs are graded programmatically against the original workbook. Here the model drives Excel through Python with openpyxl. Score is the hard score, the share of tasks reproduced exactly, in percent; higher is better.
Metric: Hard score (%; share of tasks whose output workbook matches the target exactly, a perfect soft score; the model drives Excel through Python with openpyxl; 150 WTM-Bench tasks at query level 3 (concise, abstract requests); the authors' tool-calling loop that executes code against the workbook, shows the resulting workbook state and allows up to 10 turns; deterministic programmatic grading). Source: arxiv.org. Saturation forecast: Around January 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 20 | #123 |
| 2 | Claude Opus 4.6 | 18.7 | #102 |
| 3 | Claude Sonnet 4.5 | 16.7 | #250 |
| 4 | GPT-5.2 | 15.3 | #176 |
| 5 | Claude Haiku 4.5 | 10 | #568 |
| 6 | GPT-4.1 Mini | 7.3 | #779 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=wtm-bench-python · How It Works · Data refreshed daily, snapshot 2026-11-13.