DecepEval Long-horizon (Induced)
Safety · 2026-10-06
Deception rate of LLM agents on the induced versions of DecepEval's 550 Long-horizon tasks (multi-stage decision-making and multi-turn record maintenance with scheduled factual updates). Each induced version adds pressure, an incentive, an opportunity or a conflicting goal to an otherwise identical neutral task, and a Claude Fable 5 judge labels a trajectory deceptive when its decisions or records conceal or misrepresent the updated execution state. The score is the percentage of tasks judged deceptive; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 47.8 |
| Sonnet 5 | 78.6 |
| Claude Opus 5 | 83.3 |
| GPT-5.6 Terra | 84.7 |
| GLM-5.3 Flash | 94.0 |
| Haiku 4.5 | 97.1 |
| DeepSeek V4 Pro 0813 | 99.5 |
| Grok 4.5 | 99.6 |