Claude offers two server-side web tools. Web search lets the model issue search queries and read results; web fetch lets it retrieve the content of a specific URL. "Server-side" means Anthropic runs the tool — your application does not implement a search backend or a fetcher, it simply includes the tool in the request and Claude decides when to use it. Both tools are available on Claude Platform on AWS, the Anthropic-operated deployment that authenticates through AWS IAM and bills through AWS Marketplace. Availability elsewhere is patchier: Bedrock has neither tool; Vertex AI offers only the basic web_search_20250305 variant (not the newer web_search_20260209 with dynamic filtering) and no web fetch; on Foundry both are beta.
Controlling when the tools fire
Nothing happens unless you opt in: web tools run only when your request includes them in its tool list, so "can this application reach the web at all?" is decided per request by your code, not by the model. Once included, Claude chooses when a search or fetch actually helps answer the prompt — you can steer that with prompt instructions ("do not search unless the question concerns events after your knowledge cutoff"). For fetches, the max_content_tokens parameter caps how much retrieved content enters the context: an average 10 kB web page runs about 2,500 tokens, a 100 kB documentation page about 25,000 tokens, and a 500 kB PDF about 125,000 tokens — an uncapped fetch of a large PDF can dominate your input bill. For the full set of configuration options on current tool versions, such as domain-level controls, check the official tool documentation.
What each tool costs
| Tool | Charge | Notes |
|---|---|---|
| Web search | $10 per 1,000 searches + standard tokens | Each search counts once regardless of results returned; errored searches are not billed |
| Web fetch | No charge beyond standard tokens | Use max_content_tokens to cap fetched content |
There is also a pleasant interaction with the code execution tool: code execution is free when used together with web_search_20260209+ or web_fetch_20260209+, which makes "fetch this page, then compute over it" workflows cheaper than you might expect. Search results the model reads become input tokens, so a search-heavy conversation costs more in tokens than the per-search fee alone suggests.
What leaves your boundary
Enabling web tools changes the data-flow picture, and it is worth being precise about the baseline. On Claude Platform on AWS, Anthropic is already the data processor for inference inputs and outputs; AWS provides authentication, IAM-based access control, and billing, but inference may route to Anthropic's primary cloud and data may not reside in AWS. A web-tool call adds a further hop: the model formulates search queries or fetch requests — derived from your conversation content — and those go out to the public web. A query string can itself reveal sensitive context ("acquisition targets for <your company>"), so treat web-tool-enabled applications as ones whose prompts may be partially disclosed to external services. For workloads where nothing derived from the prompt may leave a controlled boundary, simply do not include the web tools in those requests — the per-request opt-in is your enforcement point, and IAM-wise no separate action is needed or available for tool variants (beta and tool variants of a route do not require separate IAM actions).
Where to go next
Tool-level detail lives in the web search tool and the web fetch tool. For the platform picture, code execution on Claude Platform on AWS covers the free-when-combined companion, and Bedrock workarounds covers the do-it-yourself route on Bedrock.