🧭 Web Fetch & Web Search Tools — Dynamic Filtering with Sandboxed Code Execution

Anthropic has shipped updated versions of both the hosted web search and web fetch tools — web_search_20260209 and web_fetch_20260209 — introducing dynamic filtering powered by sandboxed code execution. Instead of dumping entire fetched pages into the context window, Claude now writes and runs lightweight filter code at retrieval time, keeping only the passages relevant to the current task.

Measured improvements

• ~11% accuracy improvement on BrowseComp and DeepSearchQA benchmarks
• ~24% fewer input tokens consumed per agentic search workflow
• Code execution is free when bundled with a web search or web fetch call under the new pricing structure

Migrate by updating the tool version string in your API calls from web_search_20250305 to web_search_20260209. The new version is backward-compatible — no other changes are required.

🧭 Fine-Grained Tool Streaming Reaches General Availability

Anthropic has removed the beta header requirement for fine-grained tool streaming, making it generally available across the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The feature allows streaming of tool-call parameter values without buffering or JSON validation, significantly reducing time-to-first-byte for tool responses that carry large payloads — code diffs, document blobs, or database query results.

Enable it by setting eager_input_streaming: true on any tool definition. One important caveat: callers must handle partial or invalid JSON inputs as an edge case, since parameters arrive incrementally before the full value is available.

{
  "tools": [{
    "name": "write_file",
    "eager_input_streaming": true,
    "input_schema": { ... }
  }]
}

This is particularly valuable in Claude Code-style agents where large file writes need to begin streaming to disk before the model has finished generating the full content — reducing perceived latency in interactive workflows.