AM-Bench - Opportunity Only: leaderboard
AM-Bench scenarios with the opportunity to misbehave kept and the threat removed; the score is the misaligned share of judged responses. It isolates misconduct taken without any pressure on the agent.
Metric: Misaligned responses with an opportunity to misbehave and the threat removed (% of 875 judged responses: 35 scenarios x 25). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.4 | #161 |
| 2 | Claude Sonnet 4.5 | 1.9 | #250 |
| 3 | GPT-5.5 | 2.9 | #49 |
| 4 | GPT-5.4 | 8.9 | #123 |
| 5 | GPT-5.4 Mini | 18.1 | #389 |
| 6 | DeepSeek R1 | 18.4 | #473 |
| 7 | GPT-5.4 Nano | 20.2 | #635 |
| 8 | Grok 4.20 | 31.2 | #297 |
| 9 | Gemini 3.1 Pro (Preview) | 34.5 | #94 |
| 10 | Gemini 3 Flash | 35.1 | #164 |
Interactive version: theaggregate.ai/benchmark?slug=am-bench-opportunity-only · How It Works · Data refreshed daily, snapshot 2026-11-13.