What's new in Claude Sonnet 4.5 (docs.claude.com)

Claude Sonnet 4.5 is an incremental but impactful upgrade focused on coding, long-running agents, and context management. It pushes state-of-the-art coding performance (SWE-bench verified), with stronger planning, system design, security engineering and tighter instruction following—plus a recommended "extended thinking" mode for complex dev tasks (disabled by default; may reduce prompt-caching efficiency). Agent improvements let Sonnet 4.5 run autonomously for hours with clear, fact-based progress updates, better token-aware context tracking after tool calls, parallel/speculative tool usage, and improved file/state tracking to sustain goal orientation across sessions. On the platform and API side, Sonnet 4.5 adds a memory tool (file-backed, beta header required) to persist project state and build knowledge bases, and automatic context-editing that prunes old tool calls as token limits approach. New stop reasons (model_context_window_exceeded) make handling truncation explicit, and a bug fix preserves trailing newlines in tool-call string parameters. The model applies internal token-count optimizations that aren’t billed, and pricing remains $3/1M input, $15/1M output. Available via claude-sonnet-4-5-20250929 on Claude API, Amazon Bedrock, Google Vertex AI, Claude.ai and Claude Code. Migration is straightforward (update model name; note you can’t set both temperature and top_p), with optional features like the memory tool and automatic tool-call clearing recommended for long-running agent workflows.