INS-ActBench - Spreadsheet: leaderboard
INS-ActBench (Fudan University, 2026) builds 12,050 actuarial questions from the exams of 16 actuarial associations, split into standardized knowledge, long-context insurance case analysis and spreadsheet or R-code practice tasks. This board covers the 636 spreadsheet tasks, where generated Python fills the answer cells of a workbook, scored by numerical accuracy in percent; higher is better.
Metric: Numerical accuracy (%; answer cells of the generated workbook within 0.2 percent relative tolerance) on the 636 spreadsheet tasks; 12,050 questions from the exams of 16 actuarial associations; knowledge and case questions two-shot, practice tasks zero-shot; proprietary models with thinking enabled at high where supported. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (High) | 67.7 | #49 (GPT-5.5) |
| 2 | Gemini 3.1 Pro (Preview) | 67 | #94 |
| 3 | Claude Opus 4.7 (High) | 60.6 | #79 (Claude Opus 4.7) |
| 4 | DeepSeek V4 Pro (High) | 55.2 | #188 (DeepSeek V4 Pro) |
| 5 | Qwen 3.6 Plus | 47.4 | #222 |
| 6 | Kimi K2.6 | 29.5 | #178 |
| 7 | Qwen 3.5 35B A3B | 16.3 | #482 |
| 8 | Qwen 3 14B | 2.7 | #1233 |
| 9 | Gemma 3 12B (IT) | 1 | #1837 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ins-actbench-spreadsheet · How It Works · Data refreshed daily, snapshot 2026-11-13.