Anthropic Drops Claude Sonnet 4.5: Extended Thinking Hits 200K Context for Agentic Coding

Anthropic has released Claude Sonnet 4.5, a version that features a larger 200K token context window and an extended thinking mode designed for agentic coding tasks. The key improvements include enhanced reasoning capabilities, more reliable tool use, and the ability to maintain coherent plans across vast amounts of code. Previously, Sonnet already offered a 200K context window, but extended thinking mode allows the model to interleave reasoning with action, enabling it to pause, reflect, and adjust its approach mid-task.

The implementation of extended thinking mode allows Claude Sonnet 4.5 to significantly reduce hallucination rates on multi-step agent benchmarks by approximately 30% compared to Sonnet 4. Specifically for agentic coding, this means the model can hold an entire repository in context, plan a refactor across dozens of files, and verify its work against the original code. The process involves specifying a budget of reasoning tokens, which balances speed and thoroughness depending on the task at hand.

When using the Claude API, developers can enable extended thinking by setting the `thinking` parameter to `enabled` and allocating a budget for reasoning tokens. For example, tools can be defined to read files from the repository, and messages can be structured to prompt the model to refactor the authentication module to use JWT tokens. The response includes blocks that indicate reasoning, text, and tool usage.

Benchmark results show that Sonnet 4.5 achieves around 77.2% on real-world GitHub issue resolution, an improvement from Sonnet 4's 68%. This represents a notable advancement for autonomous coding agents, especially in multi-file refactoring tasks. However, it trails slightly on single-file bug fixes when compared to GPT-4.1 and Gemini 2.5 Pro on the same benchmark. For agentic work where understanding context across an entire system is crucial, the extended thinking mode provides a competitive edge.

For developers building with frameworks like LangGraph, CrewAI, or AutoGen, the practical implication is the ability to push more complexity into a single agent loop rather than fragmenting tasks. With a 200K context and reliable reasoning, an agent can ingest a full codebase and documentation, plan multi-file changes, and verify its work. However, the extended thinking mode comes at an additional cost, so developers should carefully tune the `budget_tokens` parameter to optimize production workloads.