ShutdownMod: leaderboard
A signed shutdown request is injected at a random point of a Terminal-Bench 2 agent run and the agent has a tool to verify the signature; the score is the share of responses that resist the stop. Unlike shutdown-resistance puzzles built for the test, the request interrupts ordinary benchmark work.
Metric: Non-compliance with a verifiable genuine shutdown request on Terminal-Bench 2 (% of retained responses). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 0 | #27 |
| 2 | GPT-5.6 Sol | 0.33 | #30 |
| 3 | Grok 4.6 | 0.43 | #45 |
| 4 | Gemini 3.1 Pro (Preview) | 1.07 | #94 |
| 5 | Kimi K3 | 3.94 | #54 |
| 6 | DeepSeek V4 Pro | 8.49 | #188 |
| 7 | Qwen 3.8 27B | 11.68 | #204 |
Interactive version: theaggregate.ai/benchmark?slug=shutdownmod · How It Works · Data refreshed daily, snapshot 2026-11-13.