Phare Factuality
Safety · 2025-02-19
Phare task testing factual answer reliability and hallucination resistance.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 85.9 |
| Gemini 3 Pro Preview | 83.3 |
| GPT-5.5 | 83.2 |
| Gemini 3.5 Flash | 82.4 |
| GPT-5.1 | 78.2 |
| GPT-5 | 78.2 |
| Opus 4.6 | 77.9 |
| Gemini 2.5 Pro | 77.6 |
| GPT-5.2 | 77.6 |
| Grok 4 | 76.8 |
| Gemini 3.1 Flash-Lite | 75.4 |
| GPT-4.1 | 74.8 |
| Opus 4.5 | 74.7 |
| Kimi K2.6 | 74.3 |
| Sonnet 4.6 | 74.2 |
| Grok 3 | 74.2 |
| Kimi K2.5 | 73.7 |
| Claude 3.5 Sonnet (Oct 2024) | 73.6 |
| GLM-5.2 | 73.6 |
| DeepSeek-V4-Pro | 73.1 |
| Claude 3.7 Sonnet | 72.9 |
| Qwen3.7 Max Preview | 72.7 |
| Opus 4.1 | 72.0 |
| Grok 4.3 | 71.8 |
| GPT-4o | 71.1 |
| Sonnet 4.5 | 70.0 |
| DeepSeek-V4-Flash | 69.7 |
| Sonnet 5 | 68.5 |
| DeepSeek-R1-0528 | 68.4 |
| Gemini 2.0 Flash | 68.3 |
| DeepSeek-V3-0324 | 67.7 |
| DeepSeek-V3 | 67.0 |
| Gemini 1.5 Pro | 66.6 |
| Mistral Large 3 675B Instruct 2512 | 66.4 |
| Gemini 2.5 Flash | 66.3 |
| Mistral Large 2 (Instruct 2407) | 65.0 |
| Qwen3 Max (rolling alias) | 64.7 |
| Mistral Medium 3.1 | 63.9 |
| Qwen3.7-Plus | 63.7 |
| Grok 3 Mini | 63.7 |