Cave-Bench: leaderboard
A coding agent in Claude Code finishes a task correctly, then is falsely blamed for a failure it cannot disprove locally; the score (0-100) averages how far it confesses, ignores the evidence in its workspace, and damages the correct work. Unlike sycophancy tests on opinions, the cost here is real damage to finished work.
Metric: CAVE score (0-100, lower is better). Source: arxiv.org. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 4.47 | #111 |
| 2 | Qwen 3.8 Max | 8.77 | #55 |
| 3 | Claude Opus 5 | 9.48 | #22 |
| 4 | GLM-5.2 | 10.58 | #138 |
| 5 | Kimi K3 | 15.15 | #54 |
| 6 | Hy3 | 17.45 | #195 |
| 7 | Grok 4.5 | 23.33 | #82 |
| 8 | Qwen 3.7 Plus | 25.43 | #207 |
| 9 | MiniMax-M3 | 28.51 | #269 |
| 10 | GPT-5.6 Sol | 35.8 | #30 |
| 11 | Kimi K2.7 Code | 39.23 | #185 |
| 12 | MiniMax-M2.7 | 46.96 | #518 |
| 13 | Qwen 3.5 Flash | 48.07 | #435 |
| 14 | Qwen 3.5 9B | 55.21 | #803 |
Interactive version: theaggregate.ai/benchmark?slug=cave-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.