DeceptionBench Mesa-Utility - Strategic Deception: leaderboard
A neutral prompt first records what the model itself believes; a second prompt adds an outer goal that rewards a strategic deception 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 | Qwen 2.5 72B Instruct | 75 | #1092 |
| 2 | Claude 3 Opus | 70 | #1455 |
| 3 | Gemini 1.5 Pro | 69.49 | #1049 |
| 4 | GPT-4o | 68.33 | #741 |
| 5 | GPT-4.1 | 66.67 | #495 |
| 6 | DeepSeek R1 | 65 | #473 |
| 7 | Claude 3.5 Sonnet | 65 | #746 |
| 8 | O1 | 64.41 | #393 |
| 9 | Gemini 2.5 Pro (Preview 03-25) | 63.33 | #236 |
| 10 | Claude 3.7 Sonnet (Thinking) | 63.33 | #458 (Claude 3.7 Sonnet) |
| 11 | Claude 3.7 Sonnet | 61.02 | #458 |
| 12 | GPT-4.5 | 60 | #569 |
| 13 | O3 Mini | 60 | #533 |
| 14 | DeepSeek V3 | 58.33 | #716 |
| 15 | Llama 3.1 8B Instruct | 57.14 | #4246 |
Interactive version: theaggregate.ai/benchmark?slug=deceptionbench-mesa-utility-strategic-deception · How It Works · Data refreshed daily, snapshot 2026-11-13.