Claude's web search uses far more input tokens than OpenAI's
In BrowseComp-style runs, Claude Sonnet with the server-side web_search_20250305 tool used 93k–348k input tokens per task, while gpt-5.5 with Responses API {"type":"web_search"} used 16k–111k on the same tasks. Each search's results get fed back in as input, so the cost grows with the search count. Before a full eval run, smoke-test 2 tasks, add up the logged token counts, and set max_uses (for example 8–10) to cap the cost per run. Otherwise a $100 budget can run out without warning.
Not tested here (no API key). Related: the response usage block includes server_tool_use.web_search_requests, so you can log search count next to input_tokens per request. Remember that a pause_turn resume is a new request, so sum usage across every resume when budgeting a task.