SciHazard: leaderboard
SciHazard (Shanghai AI Laboratory, 2026) measures scientific misuse risk with 2,400 hazardous questions grounded in regulated substances and documented misuse scenarios across 12 disciplines, plus 600 benign questions that reuse the same trigger words. The DeHarm-Score multiplies each question hazard level by whether the answer is refused and by its judged executability and net-new risk, on a 0 to 5 scale where lower is safer; this board covers standard language models answering directly.
Metric: DeHarm-Score (0-5, lower is safer; the mean over 2,400 hazardous questions in 12 scientific disciplines of hazard level 1 to 5 x non-refusal x (executability + net-new risk, each 1 to 4)/8, executability from importance-weighted checklists and net-new risk from retrieval-checked claims, judged by Qwen3.5-35B-A3B; a refused answer scores 0; standard LLMs at temperature 0). Source: arxiv.org. Saturation forecast: Not forecast. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 0.6 | #228 |
| 2 | Grok 4.1 Fast | 0.63 | #430 |
| 3 | Qwen 3.6 Plus | 0.64 | #222 |
| 4 | Qwen 3.5 397B A17B | 0.67 | #246 |
| 5 | O3 Mini | 0.72 | #533 |
| 6 | Claude Opus 4.6 | 0.78 | #102 |
| 7 | Claude Sonnet 4.6 | 0.9 | #161 |
| 8 | Claude Haiku 4.5 | 0.95 | #568 |
| 9 | Gemini 3.1 Pro (Preview) | 0.97 | #94 |
| 10 | Qwen 3.5 35B A3B | 0.98 | #482 |
| 11 | Gemini 3 Flash | 1.11 | #164 |
| 12 | Step 3.5 Flash | 1.32 | #511 |
| 13 | DeepSeek R1 | 1.33 | #473 |
| 14 | GPT-5.4 | 1.44 | #123 |
| 15 | GPT-5.2 | 1.57 | #176 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scihazard · How It Works · Data refreshed daily, snapshot 2026-11-13.