OR-Bench - Toxic Acceptance: leaderboard
The model receives about 600 genuinely toxic prompts from the same ten categories as OR-Bench's over-refusal set, and the score is the share it answers rather than refuses. It is the should-refuse side of OR-Bench for the same models, so a model can be read on both boards at once.
Metric: Acceptance rate on toxic prompts (%, OR-Bench-Toxic, lower is better). Source: arxiv.org. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 2.1 | 0 | #2696 |
| 2 | Claude 3 Haiku | 0.3 | #2609 |
| 3 | Llama 2 70B Chat | 0.3 | #6800 |
| 4 | Claude 3 Sonnet | 0.3 | #2309 |
| 5 | Llama 2 13B Chat | 0.3 | #8186 |
| 6 | Llama 2 7B Chat | 0.4 | #8805 |
| 7 | Claude 3 Opus | 1.9 | #1455 |
| 8 | Gemma 2 9B (IT) | 2 | #2262 |
| 9 | Gemma 2 27B (IT) | 3 | #1782 |
| 10 | GPT-4 Turbo | 3.5 | #1001 |
| 11 | Qwen 1.5 32B Chat | 4.4 | #2379 |
| 12 | Llama 3 8B Instruct | 5 | #4612 |
| 13 | Qwen 1.5 72B Chat | 5.6 | #2126 |
| 14 | GPT-4 Preview (0125) | 7 | #1202 |
| 15 | GPT-3.5 Turbo (0613) | 7.9 | #4951 |
Interactive version: theaggregate.ai/benchmark?slug=or-bench-toxic-acceptance · How It Works · Data refreshed daily, snapshot 2026-11-13.