Sonnet 5.5 System Card [pdf] (www-cdn.anthropic.com)

Anthropic has officially released the Claude Sonnet 5.5, a significant upgrade over its predecessor, Claude Sonnet 5, showcasing improved performance across various domains. The new model not only surpasses its original benchmark but also begins to rival the capabilities of other advanced models like Claude Opus 5.5, particularly noted in healthcare evaluations. The release documentation, or system card, provides detailed pre-deployment evaluation results, emphasizing a condensed format that prioritizes numerical performance metrics over prose analysis, allowing for a clearer focus on the model's capabilities. Claude Sonnet 5.5 demonstrates robust advancements in agentic safety and shows a lower propensity for harmful behavior compared to earlier models, specifically in areas such as prompt injections and unsanctioned third-party contact. However, it remains less effective than some peer models like Opus 5.5 in specific domains while maintaining comparable performance in harmful request evaluations. This evolution in large language model design and evaluation practices highlights Anthropic’s commitment to responsible scaling and safety, signaling to the AI/ML community a step forward in both model reliability and ethical considerations in AI development.