Why does Claude search the web up to 20 times just to answer one simple question? Say you ask for the best noise-cancellation headphones under $200 released this year. First, Claude figures out what you're actually asking for. In this case, it requires real-time data. So, Claude decides to use the web search tool because Anthropic explicitly instructs it: "If you see something that requires up-to-date information, you must use the web search tool." Claude converts your prompt into something short and specific, three to four words, like best headphones 2026. Then Claude's infrastructure makes the web API call. Anthropic uses the Brave API for searching the web because Brave hands back more than just links. It gives you titles, snippets, and rich metadata. Once it has the first 10 search results, it runs a ReAct loop: reason then act on repeat. It reads what came back and picks one of three moves: • Narrow the search query and try a different search. • Open the full page of one of the search results to get more details. • Decide: "I have enough information to answer the question." Along the way, it’s constantly checking sources. A well-known review site is going to outrank a random blog. Then it loops again and again, sometimes up to 20 times, until every part of the answer is actually backed by something it found. And if we add up all the costs from the Brave API to the ReAct loop, on average, a single deep web search session costs Anthropic anywhere between 8 to 25 cents. We think of AI as a single response. Behind the scenes, it’s an army of search queries burning cash to make sure it doesn't hallucinate.