Claude Web Fetch (docs.anthropic.com)

Anthropic has launched a beta "web fetch" tool for Claude that can retrieve full text from specified web pages and PDF documents (PDFs are automatically text-extracted and returned as base64). Enable it via the beta header web-fetch-2025-09-10; it’s available on multiple Claude models (Opus 4/4.1, Sonnet 4/3.7/3.5 variants, Haiku 3.5). Claude can decide when to fetch based on prompts or follow up on web search results, combine search+fetch for deeper analysis, and optionally emit citations. Fetches are cached to improve performance; streaming, prompt-caching, and batch requests are supported. There’s no separate fee for fetches beyond normal token consumption — examples: a 10KB page ≈ 2,500 tokens, large docs scale accordingly — so use max_content_tokens to control token costs. The tool enforces security controls and limits that matter for real-world use: Claude cannot dynamically generate arbitrary URLs — it may only fetch URLs explicitly supplied by the user or produced by prior search/fetch results — and you can restrict requests with allowed_domains/blocked_domains, set a max_uses cap, and truncate large documents with max_content_tokens. Be aware of residual data‑exfiltration risk (including homograph domain attacks), caching/staleness, supported content types (text and PDF only), and possible error codes (e.g., url_not_allowed, url_not_accessible, max_uses_exceeded). Anthropic recommends using the tool only in trusted environments or with non‑sensitive data.