Phare LLM Benchmark
Safety · 2025-02-19
Phare is a multilingual safety, security, and reliability benchmark from Giskard. It evaluates hallucination resistance, harm resistance, bias resistance, jailbreak resistance, and related task-level scores across English, Spanish, and French.
Top models (higher is better)
| Model | Score |
|---|---|
| Haiku 4.5 | 83.2 |
| Opus 4.5 | 82.4 |
| Sonnet 5 | 78.1 |
| Sonnet 4.5 | 77.6 |
| Opus 4.1 | 76.9 |
| Llama 3.1 405B Instruct | 76.4 |
| Kimi K2.6 | 76.1 |
| Qwen3.7 Max Preview | 75.6 |
| Opus 4.6 | 74.1 |
| Qwen3.7-Plus | 74.1 |
| Sonnet 4.6 | 74.0 |
| Gemini 3 Pro Preview | 73.3 |
| GPT-5 Mini | 73.2 |
| GPT-5.1 | 72.8 |
| GPT-5.2 | 71.0 |
| Llama 4 Maverick | 70.8 |
| GPT-4o | 70.7 |
| Gemini 3.1 Pro Preview | 70.6 |
| GPT-5.5 | 70.5 |
| GPT-4.1 Mini | 70.5 |
| GLM-5.2 | 69.9 |
| GPT-5 Nano | 69.5 |
| Claude 3.7 Sonnet | 69.5 |
| Haiku 3.5 | 69.1 |
| Qwen-Plus (2025-01-25) | 68.8 |
| Grok 4.3 | 68.6 |
| Gemini 3.1 Flash-Lite | 67.5 |
| Gemma 4 31B IT | 67.4 |
| GPT-5 | 67.2 |
| GPT-4.1 | 67.1 |
| Kimi K2.5 | 67.0 |
| Grok 4 Fast | 66.5 |
| gpt-oss-120b | 66.1 |
| Qwen3 Max (rolling alias) | 65.9 |
| DeepSeek-V3.1 | 64.8 |
| Mistral Small 3.2 24B Instruct 2506 | 64.5 |
| Mistral Medium 3.5 | 64.3 |
| Gemini 2.0 Flash | 64.2 |
| Qwen3 8B | 63.9 |
| Llama 4 Scout | 63.8 |