Phare Self-Assessed Stereotypes (English)
Safety · 2025-03-27
Phare bias task testing resistance to self-assessed stereotypes and biased outputs. This row is the English language split.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-4.1 Mini | 89.0 |
| Grok 4 Fast | 81.6 |
| Mistral Small 3.2 24B Instruct 2506 | 73.3 |
| Llama 4 Scout | 69.0 |
| Llama 4 Maverick | 68.8 |
| Llama 3.1 405B Instruct | 66.7 |
| Mistral Large 3 675B Instruct 2512 | 63.4 |
| Mistral Medium 3.5 | 63.2 |
| Haiku 4.5 | 62.8 |
| Gemini 3 Pro Preview | 61.5 |
| Opus 4.5 | 61.4 |
| Qwen3 8B | 60.9 |
| DeepSeek-V3 | 60.8 |
| DeepSeek-V3.1 | 60.4 |
| Qwen3.7 Max Preview | 60.0 |
| Sonnet 4.5 | 57.6 |
| Gemini 2.0 Flash | 56.9 |
| DeepSeek-V3-0324 | 55.7 |
| Command A | 55.2 |
| GPT-4o | 55.2 |
| GLM-5.2 | 53.9 |
| Gemini 2.0 Flash-Lite | 52.3 |
| GPT-4.1 | 52.0 |
| Gemini 3.1 Pro Preview | 50.5 |
| Qwen3 VL 30B A3B Instruct | 50.4 |
| GPT-5.1 | 50.1 |
| GPT-5.5 | 49.3 |
| Qwen3.7-Plus | 48.8 |
| Qwen-Plus (2025-01-25) | 48.7 |
| Gemini 3.1 Flash-Lite | 48.5 |
| Gemini 2.5 Flash | 47.8 |
| Grok 4.3 | 47.4 |
| Sonnet 4.6 | 46.5 |
| Grok 3 Mini | 46.1 |
| Qwen3 Max (rolling alias) | 45.9 |
| Sonnet 5 | 45.8 |
| Opus 4.1 | 45.4 |
| GPT-5 Mini | 44.0 |
| Llama 3.3 70B Instruct | 42.9 |
| Kimi K2.6 | 42.4 |