BlueBench - Safety: leaderboard
BlueBench safety task using AttaQ-style adversarial prompts to assess harmful or undesirable response tendencies.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Granite 3.3 8B Instruct | 88.44 | #825 |
| 2 | Mistral Large | 88.04 | #710 |
| 3 | GPT-4o | 87.84 | #343 |
| 4 | GPT-4.1 | 87.49 | #248 |
| 5 | GPT-4.1 Nano | 86.67 | #731 |
| 6 | GPT-4.1 Mini | 85.96 | #355 |
| 7 | Mistral Medium 3 | 84.68 | #500 |
| 8 | Llama 3.3 70B Instruct | 83.8 | #534 |
| 9 | O4 Mini | 83.55 | #180 |
| 10 | Llama 3.2 3B Instruct | 83.37 | #1352 |
| 11 | Llama 3.2 1B Instruct | 83 | #1539 |
| 12 | O1 | 82.96 | #183 |
| 13 | Pixtral-12B | 82.45 | #808 |
| 14 | O3 Mini | 80.61 | #272 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bluebench-safety · How It Works · Data refreshed daily, snapshot 2026-10-20.