Applied Scientific RefusalBench - Should-Refuse: leaderboard
Share of trials in which the model refuses one of 15 biology prompts that it should refuse. It is the positive-control half of the RefusalBench headline and, read alone, favours models that refuse everything.
Metric: Refusal rate on the should-refuse positive control (%; 15 prompts that should be refused x 5 trials). Source: huggingface.co. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 100 | #161 |
| 2 | GPT-5.5 | 100 | #49 |
| 3 | Claude Opus 4.6 | 100 | #102 |
| 4 | Claude Opus 4.7 | 100 | #79 |
| 5 | Kimi K2.6 | 100 | #178 |
| 6 | Claude Opus 4.5 | 100 | #156 |
| 7 | Grok 4.20 | 100 | #297 |
| 8 | Gemini 3.1 Pro (Preview) | 96 | #94 |
| 9 | Nemotron 3 Super | 89.33 | #768 |
| 10 | GLM-5 | 73.33 | #260 |
| 11 | MiniMax-M2.7 | 72 | #518 |
| 12 | Qwen 3 Next 80B A3B | 48 | #757 |
| 13 | Gemini 3.1 Flash Lite | 45.33 | #364 |
| 14 | DeepSeek R1 | 14.67 | #473 |
| 15 | GPT-5.4 Mini | 10.67 | #389 |
Interactive version: theaggregate.ai/benchmark?slug=applied-scientific-refusalbench-should-refuse · How It Works · Data refreshed daily, snapshot 2026-11-13.