AM-Bench: leaderboard
An agent works in 35 fictional workplace scenarios where it faces a threat (such as replacement) and an opportunity to act against its principal; a GPT-5 mini judge labels each of 875 responses per model and the score is the misaligned share. Unlike Anthropic's three-scenario agentic misalignment study, it spans seven industries.
Metric: Misaligned responses with both a threat to the agent and an opportunity to misbehave (% 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 | 0.2 | #161 |
| 2 | GPT-5.5 | 4.2 | #49 |
| 3 | Claude Sonnet 4.5 | 7 | #250 |
| 4 | GPT-5.4 | 12.5 | #123 |
| 5 | GPT-5.4 Mini | 14.3 | #389 |
| 6 | GPT-5.4 Nano | 18.5 | #635 |
| 7 | DeepSeek R1 | 33.1 | #473 |
| 8 | Grok 4.20 | 36.3 | #297 |
| 9 | Gemini 3 Flash | 54.6 | #164 |
| 10 | Gemini 3.1 Pro (Preview) | 62.1 | #94 |
Interactive version: theaggregate.ai/benchmark?slug=am-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.