The Claude 3.5 Sonnet upgrade delivers significant improvements across benchmarks, particularly in coding and agentic tasks. It achieves 49.0% on SWE-bench Verified (up from 33.4%), outperforming all publicly available models, including specialized coding agents. It also excels in tool use, scoring 69.2% in retail and 46.0% in airline domains on TAU-bench. A major innovation is its computer use beta, enabling Claude to navigate UIs, click, type, and automate workflows—though still experimental. Early adopters like Replit and GitLab report 10% better reasoning and efficiency in multi-step coding tasks. Safety remains a priority, with joint testing by US/UK AI Safety Institutes confirming its adherence to ASL-2 risk standards.
