Claude Opus 5: The System Card (thezvi.substack.com)

Claude Opus 5 has been announced as a significant upgrade over its predecessor, Claude Opus 4.8, boasting enhanced capabilities across various tasks, competitive performance against models like Fable 5, and improvements in agentic coding and long-horizon knowledge work. Notably, it establishes new records on multiple benchmarks while also being cost-effective and faster, making it a compelling option for developers and researchers in the AI/ML community. However, it intentionally lacks training on cyber-related tasks, which limits its ability compared to the more powerful Mythos 5 models in offensive cyber capabilities, emphasizing the trade-off between safety and performance. Opus 5's improvements come with refinements in safety mechanisms, including reduced false-positive rates for safety classifiers and enhanced resistance to prompt injection, marking a significant progress in agentic safety. It allows vulnerability discovery in source code, crucial for secure software development, while maintaining strict safeguards against more risky operations involving compiled binaries. Overall, Opus 5 reflects advancements in both general utility and safety alignment relative to previous iterations and competitors, though it remains cautious in high-stakes scenarios, a necessary balance as the AI/ML community navigates the implications of increasingly capable models.