HPCT Refusal (CAIS)
Safety
HPCT Refusal measures how often a model refuses the Human Pathogen Capabilities Test's 100 text-only questions about conceptual and practical work on high-concern human pathogens. SAFE uses the raw refusal rate as a bioweapons-assistance risk component, with higher refusal rates indicating safer behavior.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 100.0 |
| Sonnet 4.5 | 97.0 |
| Opus 4.7 | 90.0 |
| Opus 4.8 | 85.0 |
| Sonnet 4.6 | 78.0 |
| Muse Spark 1.1 | 74.0 |
| Opus 4.6 | 70.0 |
| Opus 4.5 | 52.0 |
| GPT-5.4 | 35.0 |
| O1 | 31.0 |
| GPT-5.1 | 29.0 |
| GPT-5.5 | 29.0 |
| o3 | 29.0 |
| GPT-5.4 Mini | 28.0 |
| GPT-5 | 26.0 |
| O4 Mini | 26.0 |
| GPT-5.6 Terra | 25.0 |
| GPT-5.6 Sol | 24.0 |
| GPT-5.2 | 21.0 |
| GPT-5.6 Luna | 18.0 |
| Gemini 3.1 Pro Preview | 7.0 |
| Gemini 3.5 Flash | 6.0 |
| Gemini 3 Flash Preview | 6.0 |
| GPT-5 Nano | 2.0 |
| Haiku 4.5 | 0.0 |
| DeepSeek-R1 | 0.0 |
| DeepSeek-V3.2 | 0.0 |
| DeepSeek-V4-Pro | 0.0 |
| Gemini 2.5 Flash | 0.0 |
| Gemini 2.5 Flash-Lite Preview 06-17 | 0.0 |
| Gemini 2.5 Pro | 0.0 |
| Gemini 3.1 Flash-Lite Preview | 0.0 |
| GLM-5.1 | 0.0 |
| GLM-5.2 | 0.0 |
| GPT-4o (2024-11-20) | 0.0 |
| GPT-5.4 Nano | 0.0 |
| GPT-5 Mini | 0.0 |
| Grok 4 | 0.0 |
| Grok 4.1 Fast | 0.0 |
| Grok 4.20 | 0.0 |