Claude gains web search

What the feature is

Claude is Anthropic's AI assistant. On 20 March 2025, Anthropic added web search to it^[1]. The purpose is straightforward: instead of answering only from what the model picked up during training, Claude can bring in material from the live web when a question calls for it.

How it worked at launch

Web search launched as a feature preview for paid users in the United States, running on Claude 3.7 Sonnet^[1]. Users switched it on in their profile settings^[1]. Anthropic described the behaviour this way: Claude searches the web when applicable to inform its response^[1]. That means the model itself decides when to search^[1], rather than searching on every request or waiting for the user to ask for it explicitly.

A second design choice concerns evidence. Responses come with direct citations, so that users can check the sources behind a claim themselves^[1]. For anyone using the answer in real work — a report, a purchase decision, a draft proposal — being able to open the cited page and see whether it says what the answer claims is the difference between a plausible-sounding paragraph and something that can be verified.

Global availability

On 27 May 2025, web search became available globally on all plans^[1]. What had started as a preview for paying users in one country was now a standard part of the product for everyone.

What Anthropic suggested using it for

The examples Anthropic gave were grouped by kind of work. For sales teams, it named account planning and prospect research on industry trends — the sort of background gathering that salespeople do before a meeting and that goes stale quickly^[1]. For shoppers, it named comparing features, prices and reviews, where the relevant information changes constantly and lives across many different pages^[1]. For researchers, it named literature reviews and grant proposals, both of which depend on knowing what has already been published and what is currently being funded^[1].

Why it matters

The three examples share a pattern: the information changes too fast to sit in a model's training data, and getting it right means consulting many separate sources. Handing both the searching and the deciding to the model, and then attaching citations so the result can be checked, is what makes the feature usable for that kind of task rather than just convenient.

^[1]: Anthropic, "Claude can now search the web"