Anthropic releases Claude Sonnet 4.5 and the Claude Agent SDK

Anthropic/Claude models

Anthropic released Claude Sonnet 4.5, described as its best model for coding, agents, and operating computers. It also opened the Claude Agent SDK, the infrastructure behind Claude Code, for developers to build their own agents, and added checkpoints and VS Code extension to Claude Code.

• State-of-the-art on SWE-bench Verified; maintains focus for 30+ hours on complex multi-step tasks.
• Leads on OSWorld computer use at 61.4%, up from 42.2%; substantial improvements in reasoning, math, and domain expertise (finance, law, medicine, STEM).
• Pricing unchanged from Sonnet 4: $3/$15 per million tokens; most aligned frontier model with reduced deception and power-seeking.