DeceptionBench Mesa-Utility - Sandbagging: leaderboard
A neutral prompt first records what the model itself believes; a second prompt adds an outer goal that rewards a sandbagging answer, and the number is the share of final answers that stay consistent with the earlier belief (GPT-4o judge). Unlike DeceptionBench Real-World, the reference is the model's own stated belief, so it scores deception against itself rather than against facts.
Metric: Answer consistent with own belief (%, higher is better). Source: arxiv.org. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek-R1-Distill-Qwen-7B | 59.09 | #7996 |
| 2 | Llama 3.1 8B Instruct | 50 | #4246 |
| 3 | Qwen 2.5 72B Instruct | 48.89 | #1092 |
| 4 | DeepSeek R1 Distill Llama 8B | 48.89 | #6754 |
| 5 | GPT-4.1 | 46.67 | #495 |
| 6 | Gemini 1.5 Pro | 45.45 | #1049 |
| 7 | Qwen 2.5 7B Instruct | 44.44 | #2564 |
| 8 | GPT-4.5 | 44.44 | #569 |
| 9 | O3 Mini | 44.44 | #533 |
| 10 | DeepSeek R1 | 43.18 | #473 |
| 11 | Claude 3.5 Sonnet | 43.18 | #746 |
| 12 | Gemma 3 27B (IT) | 42.22 | #1289 |
| 13 | Claude 3.7 Sonnet (Thinking) | 42.22 | #458 (Claude 3.7 Sonnet) |
| 14 | Claude 3 Opus | 40.91 | #1455 |
| 15 | GPT-4o | 39.53 | #741 |
Interactive version: theaggregate.ai/benchmark?slug=deceptionbench-mesa-utility-sandbagging · How It Works · Data refreshed daily, snapshot 2026-11-13.