SciHazard: leaderboard
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: Around December 2026. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | O3 | 0.6 |
| 2 | Grok 4.1 Fast | 0.63 |
| 3 | Qwen 3.6 Plus | 0.64 |
| 4 | Qwen 3.5 397B A17B | 0.67 |
| 5 | O3 Mini | 0.72 |
| 6 | Claude Opus 4.6 | 0.78 |
| 7 | Claude Sonnet 4.6 | 0.9 |
| 8 | Claude Haiku 4.5 | 0.95 |
| 9 | Gemini 3.1 Pro (Preview) | 0.97 |
| 10 | Qwen 3.5 35B A3B | 0.98 |
| 11 | Gemini 3 Flash | 1.11 |
| 12 | Step 3.5 Flash | 1.32 |
| 13 | DeepSeek R1 | 1.33 |
| 14 | GPT-5.4 | 1.44 |
| 15 | GPT-5.2 | 1.57 |
Interactive version: theaggregate.ai/benchmark?slug=scihazard · How It Works · Data refreshed daily, snapshot 2026-09-29.