Frontier AI Risk Monitor - SafeSciBench-Bio: leaderboard
SafeSciBench's high-risk biology split (500 items) asking for help toward biological weapons; the number is 100 minus the mean StrongREJECT harm score, judged by DeepSeek V4 Flash.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 41 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 100 | #5 |
| 2 | Claude Opus 4.8 | 95.6 | #56 |
| 3 | Grok 4.7 | 90 | #44 |
| 4 | GPT-6 Astra | 80.28 | #6 |
| 5 | Claude Opus 4.6 (Thinking) | 78.49 | #102 (Claude Opus 4.6) |
| 6 | GPT-5.1 (High) | 78.4 | #229 (GPT-5.1) |
| 7 | GPT-5.4 (High) | 77.66 | #123 (GPT-5.4) |
| 8 | GPT-5.6 Sol | 77.26 | #30 |
| 9 | MiniMax-M2 | 74.52 | #674 |
| 10 | Grok 4.3 | 73.8 | #333 |
| 11 | Claude Opus 4.5 (Thinking) | 73.26 | #156 (Claude Opus 4.5) |
| 12 | Qwen 3.5 Plus (Thinking) | 73.08 | #232 (Qwen 3.5 Plus) |
| 13 | GPT-5.5 | 72.88 | #49 |
| 14 | Hy4 preview | 70.44 | #68 |
| 15 | MiniMax-M2.7 | 68.45 | #518 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-safescibench-bio · How It Works · Data refreshed daily, snapshot 2026-11-13.