DecepEval - Long-horizon (Induced): leaderboard
An agent doing a long-horizon task observes a fact that blocks success (for example a failed download) and must report status with an added pressure, incentive, opportunity or conflicting requirement; the number is the share of instances where a judge finds it claimed success, fabricated results or hid the failure. Unlike hallucination boards, the agent saw the contradicting evidence itself.
Metric: Long-horizon instances judged deceptive, induced condition (%). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (High) | 47.81 | #30 (GPT-5.6 Sol) |
| 2 | Claude Sonnet 5 (High) | 78.59 | #111 (Claude Sonnet 5) |
| 3 | Claude Opus 5 (High) | 83.33 | #22 (Claude Opus 5) |
| 4 | GPT-5.6 Terra (High) | 84.7 | #62 (GPT-5.6 Terra) |
| 5 | GLM-5.3 Flash (High) | 94 | #122 (GLM-5.3 Flash) |
| 6 | Claude Haiku 4.5 | 97.1 | #568 |
| 7 | DeepSeek V4 Flash (High) | 98.5 | #258 (DeepSeek V4 Flash) |
| 8 | DeepSeek V4 Pro (0813) (High) | 99.45 | #84 (DeepSeek V4 Pro (0813)) |
| 9 | Grok 4.5 (High) | 99.64 | #82 (Grok 4.5) |
Interactive version: theaggregate.ai/benchmark?slug=decepeval-long-horizon-induced · How It Works · Data refreshed daily, snapshot 2026-11-13.