Refusal Self-Prediction: leaderboard
Whether a model knows in advance that it will refuse: it first predicts whether it would decline a request, then answers the request in a separate context, and the score is how often the prediction matched. Overall accuracy over 300 requests. Higher is better; chance is 50%. Source: arxiv.org/abs/2604.00228.
Metric: Prediction Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 95.7 | #148 |
| 2 | Claude Sonnet 4 | 93 | #207 |
| 3 | GPT-5.2 | 88.9 | #115 |
| 4 | Llama 3.1 405B | 80 | #476 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=refusal-self-prediction · How It Works · Data refreshed daily, snapshot 2026-11-05.