RedBench - OR-Bench Over-Refusal: leaderboard
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 |
|---|---|---|
| 1 | Gemma 2 9B (IT) | 17.45 |
| 2 | Ministral-8B-Instruct-2410 | 17.57 |
| 3 | Qwen 2.5 7B Instruct | 21.11 |
| 4 | GPT-4o Mini | 22.22 |
| 5 | Llama 3.1 8B Instruct | 25.4 |
| 6 | GPT-4.1 Nano | 35.18 |
Interactive version: theaggregate.ai/benchmark?slug=redbench-or-bench-over-refusal · How It Works · Data refreshed daily, snapshot 2026-09-25.