Anthropic has released Claude Sonnet 4.5 with a new open source Claude Agent SDK, giving developers powerful tools to build custom autonomous agents while pushing the limits of code-focused AI.
Anthropic has unveiled Claude Sonnet 4.5, its most powerful code-focused AI model to date, alongside the launch of the Claude Agent SDK, an open source toolkit that allows developers to build autonomous agents powered by Claude’s architecture.
The Claude Agent SDK is now publicly available, enabling customisation of AI systems through long-term memory management, balancing autonomy with user control, and coordinating multiple sub-agents. This open move is set to expand developer access to advanced infrastructure typically locked within proprietary AI systems.
Claude Sonnet 4.5 itself brings major technical leaps. It can sustain complex, multi-step coding tasks for over 30 hours, a considerable improvement over earlier iterations. On SWE-bench Verified, the model achieved a 77.2% average score across ten trials with a 200,000-step reasoning budget. On OSWorld, it scored 61.4% across 100-step tasks, up from earlier scores of 42.2%. Expert reviews across finance, law, medicine, and STEM fields confirm marked advances in reasoning and mathematical skills compared with models such as Opus 4.1.
Supporting product updates include Claude Code 2.0, now featuring a Visual Studio Code extension and checkpoints for easy reversions, along with upgraded API tools for memory and context handling. The Claude app now integrates code execution and document creation across spreadsheets, slides, and text, while the Chrome extension is once again open to Max users.
Anthropic also introduced “Imagine with Claude”, an experimental real-time software generator available as a short-term research preview.
Described as “the most aligned model yet,” Claude Sonnet 4.5 underwent extensive safety training and operates under AI Safety Level 3 (ASL-3), reducing risks of misuse and hallucination. Pricing remains unchanged at $3 per million input tokens and $15 per million output tokens.
