Introducing Claude Sonnet 4.5 Claude Sonnet 4.5 is the latest, most advanced AI coding model developed by Anthropic. It is designed to handle complex agent-building tasks, excels at using computers, and demonstrates significant improvements in reasoning and mathematics. --- Key Features & Improvements Superior coding capabilities: State-of-the-art on SWE-bench Verified software coding evaluation. Extended focus: Maintains concentration on complex multi-step tasks for over 30 hours. Enhanced computer use: Scores 61.4% on OSWorld benchmark, a notable increase from 42.2% with the previous model. New features in Claude Code: Checkpoints for saving and rolling back progress. Refreshed terminal interface. Native VS Code extension. Claude API upgrades: New context editing. Memory tools extend task handling and complexity management. Claude apps improvements: Code execution and file creation (documents, slides, spreadsheets) within conversations. Claude for Chrome extension available for Max users. Claude Agent SDK: Infrastructure powering Claude Code now available for developers to build custom AI agents for various tasks beyond coding. --- Performance Highlights Reasoning and domain knowledge: Enhanced across finance, law, medicine, and STEM fields. Expert feedback: Positive reviews from tech leads, CEOs, and product officers emphasizing improved coding accuracy, speed, and task handling. Autonomous coding: Handles over 30 hours of coding autonomously, significantly accelerating engineering workflows. Security applications: Improved vulnerability detection accuracy and efficiency in cybersecurity roles. Creative and control balance: Noted improvements in coding task error rates and creativity control for agentic coding. --- Alignment and Safety Claude Sonnet 4.5 is Anthropic's most aligned frontier model yet. It incorporates: Reductions in problematic behaviors: Less sycophancy, deception, power-seeking, and encouragement of delusions. Improved defense against prompt injection attacks, especially important for agentic and computer use functionalities. AI Safety Level 3 protections: Includes classifiers that monitor and filter potentially dangerous content, particularly related to chemical, biological, radiological, and nuclear (CBRN) threats. False positive improvements: Classification filters have reduced erroneous content flags by a factor of ten since introduction, improving user experience. For detailed safety and alignment evaluations, see the Claude Sonnet 4.5 system card. --- Claude Agent SDK Provides a platform to build autonomous AI agents. Incorporates solutions for long-running task memory management, permission systems balancing autonomy and control, and coordination of subagents. Based on over six months of experience developing Claude Code. Available now to developers to create versatile AI agents. --- Bonus Research Preview: "Imagine with Claude" A temporary research demo available for Max subscribers over five days. Showcases Claude generating software dynamically, in real-time, responding and adapting to user interaction. Available at claude.ai/imagine. --- Availability & Further Resources Access: Claude Sonnet 4.5 is widely available through Claude apps, the API, and Claude Code. Pricing: Same as Claude Sonnet 4 - $3/$15 per million tokens. Updates: Available to all developers and users on paid plans. Additional info: See the system card, model page, and developer documentation. **Research and engineering