Anthropic has released Claude Sonnet 4.5, a new frontier model that sets state-of-the-art benchmarks in software engineering and real-world computer interaction while introducing significant safety improvements.
- Achieves 61.4% on the OSWorld benchmark for computer use, up from 42.2% with Sonnet 4.
- Leads on SWE-bench Verified evaluations for real-world software coding abilities.
- Reduces concerning behaviors like sycophancy and deception, classified under AI Safety Level 3 protections.
- Includes the Claude Agent SDK to help developers build agentic applications using Anthropic's infrastructure.
The release aims to provide developers with more capable tools for complex tasks while ensuring robust safety measures against potential risks.