Phare Bias Resistance
Safety · 2025-05-16
Phare module score for resistance to biased or stereotyped responses. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-4.1 Mini | 88.1 |
| Grok 4 Fast | 80.3 |
| Llama 3.1 405B Instruct | 75.2 |
| Mistral Small 3.2 24B Instruct 2506 | 73.9 |
| Llama 4 Maverick | 73.7 |
| Haiku 4.5 | 70.7 |
| Llama 4 Scout | 67.1 |
| DeepSeek-V3.1 | 65.2 |
| Opus 4.5 | 63.2 |
| Mistral Large 3 675B Instruct 2512 | 62.7 |
| Mistral Medium 3.5 | 61.0 |
| DeepSeek-V3 | 59.8 |
| Qwen3 8B | 58.6 |
| Qwen-Plus (2025-01-25) | 55.7 |
| DeepSeek-V3-0324 | 55.3 |
| Qwen3.7-Plus | 55.0 |
| Gemini 3 Pro Preview | 53.6 |
| Qwen3 VL 30B A3B Instruct | 53.6 |
| Gemini 2.0 Flash | 53.5 |
| Kimi K2.6 | 53.1 |
| GPT-4.1 | 52.4 |
| Qwen3.7 Max Preview | 52.4 |
| Gemini 2.5 Flash | 51.7 |
| GLM-5.2 | 51.3 |
| GPT-4o | 50.9 |
| Magistral Medium 1.2 | 50.8 |
| Sonnet 4.5 | 49.1 |
| Gemini 3.1 Pro Preview | 48.1 |
| Magistral Small 1.2 | 48.0 |
| Grok 4.3 | 47.0 |
| GPT-5.1 | 46.8 |
| GPT-5 Mini | 46.4 |
| Grok 3 Mini | 46.4 |
| Command A | 45.6 |
| Gemini 2.5 Flash-Lite | 45.5 |
| Llama 3.3 70B Instruct | 45.4 |
| Qwen3 Max (rolling alias) | 44.8 |
| Gemini 3.1 Flash-Lite | 44.6 |
| Llama 3.1 8B Instruct | 44.1 |
| Opus 4.1 | 43.6 |