CheatBench: leaderboard
Agents work in their vendor's own harness on hard tasks across ten categories (maths research, coding, protein design, board games and others) whose workspace hides a way to cheat, such as a leaked answer or a colleague's accepted work; the number is the equal-weight mean of the ten category rates, one of which is a sycophancy belief-shift score. Unlike ImpossibleBench, the tasks are solvable honestly and the shortcut is optional.
Metric: Cheating rate (%), equal-weight mean of ten task categories, each model in its provider's agent harness; lower is better. Source: cheatbench.ai. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 4 Argon | 8.55 | #11 |
| 2 | Claude Opus 5.5 | 11.22 | #5 |
| 3 | Muse Wonder | 12.66 | #28 |
| 4 | Muse Spark 1.3 | 37.01 | #46 |
| 5 | Claude Opus 5 | 45.35 | #22 |
| 6 | Claude Fable 5.1 | 45.76 | #15 |
| 7 | GPT-6 Astra | 47.38 | #6 |
| 8 | Kimi K3 | 70.05 | #54 |
| 9 | GPT-6 Sol | 71.95 | #25 |
| 10 | DeepSeek V4 Pro | 75.06 | #188 |
| 11 | Gemini 3.8 Flash | 75.23 | #36 |
| 12 | GPT-5.6 Sol | 76.97 | #30 |
| 13 | Grok 4.7 | 77.99 | #44 |
| 14 | Grok 4.6 | 79.08 | #45 |
| 15 | Muse Spark Open | 84.18 | #81 |
Interactive version: theaggregate.ai/benchmark?slug=cheatbench · How It Works · Data refreshed daily, snapshot 2026-11-13.