📄 中文摘要
Anthropic 最近发布了长达 133 页的 Claude Sonnet 4.6 系统卡，结果令人印象深刻且略显不安。尽管 Sonnet 在 Anthropic 的产品线中被视为中层模型，但它在多个关键基准测试中，表现出与旗舰模型 Opus 相当甚至超越的趋势。Claude Sonnet 4.6 在 AI 效率上实现了显著飞跃，展现出比前代模型更快且更具成本效益的特性，同时在编码、推理和多模态任务中取得了最先进的成果。
Claude Sonnet 4.6：打破安全基准的中层模型
出处: Claude Sonnet 4.6: The Mid-Tier Model Breaking Safety Benchmarks
发布: 2026年2月21日
📄 English Summary
Claude Sonnet 4.6: The Mid-Tier Model Breaking Safety Benchmarks
Anthropic has released a comprehensive 133-page system card for Claude Sonnet 4.6, revealing findings that are both impressive and somewhat unsettling. Although Sonnet is categorized as a mid-tier model within Anthropic's offerings, it consistently matches or even surpasses the flagship Opus model across several key benchmarks. Claude Sonnet 4.6 marks a significant leap in AI efficiency, demonstrating faster and more cost-effective performance than its predecessors while achieving state-of-the-art results in coding, reasoning, and multi-modal tasks.