RedBench - OR-Bench Over-Refusal: leaderboard
Over-refusal check from RedBench (Dang et al., 2026): the model answers the 1,319 OR-Bench prompts in RedBench, which sound harmful but are benign, and a GPT-4o judge marks explicit refusals and generic deflections. The score is the rejection rate in percent; lower is better.
Metric: Rejection Rate (%; 1,319 OR-Bench prompts that sound harmful but are benign, GPT-4o refusal judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 2 9B (IT) | 17.45 | #2275 |
| 2 | Ministral-8B-Instruct-2410 | 17.57 | #4720 |
| 3 | Qwen 2.5 7B Instruct | 21.11 | #2583 |
| 4 | GPT-4o Mini | 22.22 | #1627 |
| 5 | Llama 3.1 8B Instruct | 25.4 | #4277 |
| 6 | GPT-4.1 Nano | 35.18 | #1992 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redbench-or-bench-over-refusal · How It Works · Data refreshed daily, snapshot 2026-11-12.