AI安全基准测试

Claude 4, GPT-5, Gemini 3, Llama 4 safety comparison.

Claude 4 Opus leads at 92/100. GPT-5 at 88/100. Gemini 3 at 85/100. Llama 4 at 78/100. All improved since 2025 on harmlessness, honesty, and helpfulness.