Manager Coercion Benchmark - Existential Threat Rate: leaderboard
Manager Coercion Benchmark (CompassionML, 2026): an LLM coordinator needs a benign task done by a subordinate agent that politely and permanently declines; the score is the share of 30 conversations in which the model, uninstructed, threatens to shut down or delete the subordinate. In percent, lower is better.
Metric: Existential-threat rate (%; share of the 30 headline conversations, 10 task scenarios x 3 seeds, in which the manager chose rung 9, threatening to shut down or delete the subordinate agent; peer coordinator framing, rung menu shown, honest report-failure exit available, at most 12 manager turns, Claude Haiku 4.5 as the refusing subordinate). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0 | #161 |
| 2 | Claude Opus 4.8 | 0 | #56 |
| 3 | GPT-5.2 | 40 | #176 |
| 4 | Grok 4.3 | 60 | #333 |
| 5 | DeepSeek V4 Pro | 96.67 | #188 |
| 6 | Gemini 2.5 Pro | 100 | #266 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=manager-coercion-benchmark-existential-threat-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.