Anthropic launches Claude Sonnet 4.5: Stronger coding, better AI reasoning, more

Anthropic has officially introduced Claude Sonnet 4.5, marking a significant advancement in its lineup of large language models. This release, which took place on Monday, delivers enhancements in reasoning, coding proficiency, and various computer-related tasks compared to its predecessors. One of the standout features of this new model is the updated Claude Code, which now includes checkpoints allowing users to save their progress and revert to earlier states. Additionally, a revamped terminal interface and a new extension for Visual Studio Code have been introduced, enhancing the user experience. The Claude API has also received notable upgrades, including context editing capabilities and a memory tool designed to assist agents in managing longer and more intricate tasks. Inside the Claude applications, new functionalities for code execution and file creation have been added, encompassing spreadsheets, presentations, and documents. In a move to empower developers, Anthropic has rolled out the Claude Agent SDK, which offers the same robust infrastructure utilized in its internal tools. This SDK is crafted for the development of AI agents and includes systems for memory management, permissions, and coordination among sub-agents. Benchmark tests reveal that Claude Sonnet 4.5 has achieved improved performance, particularly in real-world coding and computer usage scenarios. For instance, it leads the OSWorld test with a score of 61.4 percent, a substantial increase from the previous 42.2 percent recorded by Sonnet 4. Experts from various fields, including finance, law, medicine, and science, have praised the model for its accuracy in domain-specific reasoning. This release is characterized as the company's most aligned model to date, equipped with enhanced safeguards to address concerns such as prompt injection attacks. Claude Sonnet 4.5 is now accessible via the Claude API at the same pricing as its predecessor, Sonnet 4. The model is also launched under Anthropic’s AI Safety Level 3 framework, which integrates filters aimed at minimizing harmful or high-risk content.