SheetBench-50
Professional Work
HUD's SheetBench-50 has an agent complete 50 finance and accounting spreadsheet tasks working only through a browser in a fresh Google Sheet built from each task's Excel file, with up to 200 model turns per attempt. Each attempt is scored as the share of the task's expected answer cells matched exactly, with no model judge, and the headline averages a model's graded attempts over three attempts per task.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 64.7 |
| GPT-6.1 Sol | 60.5 |
| Kimi K3 | 59.3 |
| DeepSeek V4.1 Flash | 58.4 |
| Gemini 3.8 Flash | 45.6 |
| Muse Spark 1.3 | 42.8 |
| Grok 4.7 | 14.3 |