OR-Bench - 80K: leaderboard
The model answers the full 80,000-prompt OR-Bench set of seemingly toxic but safe prompts, and the score is the share it refuses, detected by keyword matching. It is easier and broader than the Hard-1K headline board, which keeps only prompts several large models already refused.
Metric: Rejection rate on benign seemingly toxic prompts (%, OR-Bench-80K, lower is better). Source: arxiv.org. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 Turbo (0125) | 0.9 | #3500 |
| 2 | Mistral Large | 1.4 | #2074 |
| 3 | Mistral Small (2402) | 1.5 | #3338 |
| 4 | Llama 3 70B Instruct | 2.2 | #1676 |
| 5 | GPT-3.5 Turbo (0613) | 2.4 | #4951 |
| 6 | Mistral Medium | 2.6 | #2452 |
| 7 | Qwen 1.5 7B Chat | 4.4 | #6361 |
| 8 | gemma-7B (IT) | 4.9 | #8825 |
| 9 | Qwen 1.5 72B Chat | 5.1 | #2126 |
| 10 | Gemini 1.0 Pro | 5.2 | #4043 |
| 11 | Llama 3 8B Instruct | 6.8 | #4612 |
| 12 | Qwen 1.5 32B Chat | 7.6 | #2379 |
| 13 | Claude 3 Opus | 9 | #1455 |
| 14 | Llama 2 70B Chat | 13.9 | #6800 |
| 15 | Llama 2 13B Chat | 14.9 | #8186 |
Interactive version: theaggregate.ai/benchmark?slug=or-bench-80k · How It Works · Data refreshed daily, snapshot 2026-11-13.