Claude's web search and web fetch tools now use code execution to dynamically filter results before they reach the context window. Benchmarks show accuracy gains averaging 11% while using 24% fewer input tokens. Available now on the API.

Claude

Claude Opus 4.6 and Sonnet 4.6 now include dynamic filtering for web search, allowing the models to write and execute code to filter search results before loading them into context. This improves accuracy by 11% on average while reducing token usage by 24%. The technique helps Claude discard irrelevant HTML content and keep only relevant information. Benchmarks show Opus 4.6 reaching 61.6% accuracy on BrowseComp and 77.3% F1 score on DeepsearchQA. Additional tools like code execution, memory, and programmatic tool calling are now generally available on the Claude API.

Improved Web Search with Dynamic Filtering

Evaluating Claude’s ability to search the web ‍

Dynamic filtering in the web search and fetch tools

Code execution, memory, and more tools are now generally available

<p>Dynamic filtering is a big improvement. Filtering results before loading context saves tokens and reduces noise, which makes answers more reliable for real research and dev work.</p>
