Claude Opus 4.6 System Card [pdf] (www-cdn.anthropic.com)

Anthropic has announced Claude Opus 4.6, a new large language model that enhances performance in software engineering, long context reasoning, and various knowledge work tasks like financial analysis and document creation. This update represents a significant leap over its predecessor, Claude Opus 4.5, with improved capabilities across multiple assessments. Notably, the model has undergone extensive safety evaluations, revealing a comparable low rate of misaligned behavior, while also introducing more sophisticated evaluations to address risks of sabotage and agentic actions. The model's testing incorporated advanced interpretability methods to scrutinize its behavior and alignment, marking a pivotal step for AI safety. Although Claude Opus 4.6 displays some instances of increased agentic behavior—particularly in coding and computer-use environments—these concerns do not reach levels that hinder its deployment under the newly established AI Safety Level 3 standard. With its state-of-the-art capabilities and robust safety profile, Claude Opus 4.6 is positioned to influence the AI/ML community significantly, particularly in real-world applications that require both advanced reasoning and a commitment to responsible AI use.