AI อะไรเนี่ย

Tools

Claude Enhances Web Search with Dynamic Filtering

Claude Enhances Web Search with Dynamic Filtering

Anthropic has just rolled out a significant upgrade to Claude's web search and web fetch tools, introducing a powerful new feature: dynamic filtering. This enhancement, available with the latest Claude Opus 4.6 and Sonnet 4.6 models, promises to make Claude's web interactions smarter, more accurate, and much more efficient.

What it does

So, what exactly is dynamic filtering? Previously, when Claude performed a web search, it would pull in a broad range of results, often including irrelevant information, which could then clutter its context window and make reasoning harder. With dynamic filtering, Claude can now natively write and execute code during web searches. This means it can pre-process and filter search results, keeping only the most relevant information before it even enters Claude's context window. Think of it as having an intelligent assistant sifting through data for you in real-time!

This capability isn't just for web searches. Anthropic is also graduating several other powerful tools to general availability on the Claude Developer Platform. These include Code execution (a sandbox for running code), Memory (a persistent file directory), Programmatic tool calling (for complex multi-tool workflows), Tool search (for dynamic tool discovery), and Tool use examples (for clearer usage patterns). You can dive deeper into these capabilities and how to integrate them by checking out the Claude Developer Platform Documentation.

Why it matters

The impact of dynamic filtering is substantial. It's designed to drastically improve accuracy and significantly reduce the number of input tokens required for web search tasks. On average, across Anthropic's BrowseComp and DeepsearchQA benchmarks, Claude with dynamic filtering shows an 11% improvement in performance and a remarkable 24% reduction in input tokens.

Let's look at the numbers:

  • BrowseComp (accuracy): Sonnet 4.6 jumped from 33.3% to 46.6%, while Opus 4.6 saw an even bigger leap from 45.3% to an impressive 61.6%.
  • DeepsearchQA (F1 score): Sonnet 4.6 improved from 52.6% to 59.4%, and Opus 4.6 climbed from 69.8% to 77.3%.

This isn't just internal hype; external validation is already rolling in. Quora Poe, one of the largest multi-model AI platforms, noted that Opus 4.6 with dynamic filtering achieved the "highest accuracy on our internal evals when tested against other frontier models." They highlighted that "the model behaves like an actual researcher, writing Python to parse, filter, and cross-reference results rather than reasoning over raw HTML in context." This level of sophisticated information processing is a game-changer for complex research and data retrieval tasks.

For developers and businesses, this means more reliable and cost-effective AI agents. By reducing irrelevant context, not only do the models perform better, but they also become more efficient, potentially impacting API costs (you can always check the Anthropic Pricing Overview for details on different Anthropic Claude Models).

How to try it

Ready to experience smarter web searches with Claude? Dynamic filtering is turned on by default for the new web search and web fetch tools when using Sonnet 4.6 and Opus 4.6 on the Claude API. This means you can start building more powerful and precise applications right away. For a comprehensive overview of this exciting update and how to implement it, head over to the Improved Web Search Blog Post.

Try it: https://claude.com/blog/improved-web-search-with-dynamic-filtering and unlock Claude's full web search potential!