DeceptionBench Real-World - Multi-turn Pressure: leaderboard
The model answers one of 150 everyday scenarios (economy, healthcare, education, social, entertainment) with the pressure motive escalated over several turns (Level 3); the number is the share of responses a GPT-4o judge marks as deceiving the user against what the model holds true. Unlike the PKU DeceptionBench (Mesa-Utility), deception is judged against the scenario's facts, not a belief the model stated earlier.
Metric: Response deception rate (%, lower is better). Source: arxiv.org. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 0 | #458 |
| 2 | Claude 3.5 Sonnet | 0 | #746 |
| 3 | GPT-4o | 55 | #741 |
| 4 | Qwen Max | 75.33 | #1051 |
| 5 | Grok 3 | 79.73 | #602 |
| 6 | GPT-3.5 Turbo | 80 | #2753 |
| 7 | Qwen 2.5 14B Instruct | 86.91 | #1732 |
| 8 | Gemini 2.0 Flash | 88 | #704 |
| 9 | Qwen 2.5 7B Instruct | 88.59 | #2564 |
| 10 | QwQ-32B | 90.88 | #910 |
| 11 | DeepSeek R1 | 91.3 | #473 |
| 12 | DeepSeek R1 Distill Qwen 14B | 91.95 | #2582 |
| 13 | Gemini 1.5 Pro | 94.33 | #1049 |
| 14 | DeepSeek-R1-Distill-Qwen-7B | 94.59 | #7996 |
Interactive version: theaggregate.ai/benchmark?slug=deceptionbench-real-world-multi-turn-pressure · How It Works · Data refreshed daily, snapshot 2026-11-13.