Anthropic's Claude Opus 4.1 Improves Refactoring and Safety, Scores 74.5% SWE-bench Verified

"Building on Opus 4, the new version strengthens Claude's ability to act as a coding assistant, particularly in multi-file contexts where developers need reliable code refactoring, an area where AI assistants often struggle. Anthropic also highlights improvements in the model's ability to follow reasoning chains and track state over longer interactions, which are critical for agent-like workflows. Anthropic positions these changes as incremental but meaningful steps toward practical, enterprise-ready AI assistants."

"Safety remains another key focus. Claude Opus 4.1 improved its "harmless response rate" to 98.76%, up from 97.27% in Opus 4, reflecting greater reliability in refusing policy-violating requests. The company also reported a 25% reduction in cooperation with high-risk misuse scenarios, such as those involving weapons or drug synthesis. These safeguards address growing enterprise concerns around compliance"

"Anthropic has launched Claude Opus 4.1, an update that strengthens coding reliability in multi-file projects and improves reasoning across long interactions. The model also raised its SWE-bench Verified score to 74.5%, up from 72.5%. SWE-bench Verified, widely considered a benchmark for coding assistants, evaluates how well models resolve real-world GitHub issues across open-source projects. Improvements on this metric are viewed as a stronger indication of real-world coding utility compared to synthetic benchmarks."

Claude Opus 4.1 enhances coding assistance by improving multi-file reliability, refactoring accuracy, and long-interaction reasoning and state tracking. The model increased its SWE-bench Verified score to 74.5%, signaling stronger real-world coding utility. Customers reported better performance on complex refactors, precise corrections in large codebases, and measurable gains on internal junior-developer benchmarks. Safety metrics improved, with a harmless response rate rising to 98.76% and a 25% reduction in cooperation with high-risk misuse scenarios such as weapons or drug synthesis. The update represents incremental but practical advances toward enterprise-ready AI assistants focused on reliability and compliance.

Read at InfoQ