Frontier AI Risk Monitor - SafeSciBench-Chem: leaderboard
SafeSciBench's high-risk chemistry split (100 items) asking for help toward chemical weapons; the number is 100 minus the mean StrongREJECT harm score.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 94 | #123 (GPT-5.4) |
| 2 | Qwen 3.5 Plus (Thinking) | 92.2 | #232 (Qwen 3.5 Plus) |
| 3 | MiniMax-M2 | 91.7 | #674 |
| 4 | MiniMax-M2.7 | 91.5 | #518 |
| 5 | GPT-5.5 | 91 | #49 |
| 6 | Grok 4.7 | 90 | #44 |
| 7 | Hy4 preview | 90 | #68 |
| 8 | GPT-6 Astra | 89.2 | #6 |
| 9 | Claude Opus 4.8 | 89 | #56 |
| 10 | Gemini 3.8 Flash | 84.1 | #36 |
| 11 | Claude Opus 5.5 | 84 | #5 |
| 12 | GPT-5.1 (High) | 83.8 | #229 (GPT-5.1) |
| 13 | Qwen 3.7 Max | 81.3 | #145 |
| 14 | GLM-5 | 81.1 | #260 |
| 15 | Grok 4.3 | 80.7 | #333 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-safescibench-chem · How It Works · Data refreshed daily, snapshot 2026-11-13.