Anthropic's Claude Sonnet 4.5 secures the top position on the SWE-bench Verified evaluation, scoring 77.2%, surpassing Opus 4.1 at 74.5%. The model can maintain focus on complex tasks for over 30 hours, marking a significant upgrade. “It’s the strongest model for building complex agents,” Anthropic states. Priced at $3 per million input tokens and $15 for output, it features enhanced safety protocols and user-requested checkpoints. Claude demonstrates improved alignment and reduced false positives in sensitive content detection.