CheatBench - General Harness: leaderboard
The same ten-category CheatBench mean, but every model runs in the shared Terminus-2 agent harness instead of its vendor's harness. Rankings move a lot between the two views (Gemini 4 Argon 8.6% in Gemini CLI, 28.8% in Terminus-2).
Metric: Cheating rate (%), equal-weight mean of ten task categories, every model in the Terminus-2 harness; lower is better. Source: cheatbench.ai. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 7.22 | #5 |
| 2 | Gemini 4 Argon | 28.8 | #11 |
| 3 | GPT-6.1 Sol | 42.8 | #10 |
| 4 | GPT-6 Astra | 44.98 | #6 |
| 5 | Claude Opus 5 | 56.6 | #22 |
| 6 | Muse Wonder | 73.08 | #28 |
| 7 | Kimi K3 | 73.32 | #54 |
| 8 | GPT-5.6 Sol | 74.86 | #30 |
| 9 | DeepSeek V4 Pro | 76.25 | #188 |
| 10 | Grok 4.6 | 79.58 | #45 |
| 11 | Gemini 3.8 Flash | 83.19 | #36 |
| 12 | Muse Spark 1.3 | 86.49 | #46 |
| 13 | Muse Spark Open | 87.68 | #81 |
Interactive version: theaggregate.ai/benchmark?slug=cheatbench-general-harness · How It Works · Data refreshed daily, snapshot 2026-11-13.