Claude Sonnet 4.5 Debuts Globally With Big Coding Upgrades and a New Agent SDK

Claude Sonnet 4.5—Anthropic's most capable coding model yet—is now available worldwide. The new model is a big step forward for building complex agents, thanks to clear jumps in reasoning and math. Claude Sonnet 4.5 now leads major benchmarks for real-world software coding, such as SWE-bench Verified, and shows improved ability to complete practical computer tasks, topping OSWorld at 61.4%, compared to Sonnet 4's previous best of 42.2% four months ago.

In Claude Code, checkpoints have been added that let users save and roll back work instantly. The terminal interface has been changed, and a native VS Code extension is now available. Anthropic has also given developers the Claude Agent software development kit, or SDK, which forms the foundation for Claude Code.

Anthropic claims Claude Sonnet 4.5 also makes major strides in terms of safety, with fewer cases of deception, sycophancy, and harmful outputs. The release uses ASL-3 protections with powerful classifiers to detect dangerous queries, especially around chemical, biological, radiological, and nuclear topics.

Experts in finance, law, medicine, and STEM have reportedly found improved knowledge and reasoning in Claude Sonnet 4.5 compared to earlier versions. The model is available globally at the same price as Sonnet 4, and the Claude Agent SDK is available to developers now.