Anthropic ran a modified harness with a 980k-token task budget, offline web search and fetch tools limited to a frozen index, and Claude Opus 4.6 as grader; results are not directly comparable with Perplexity's official live-web, GPT-5.4-graded numbers.Perplexity's wide-search research benchmark (Wide ANd Deep Research): 500 tasks require finding tens to hundreds of entities and facts about each, cited in structured records; the primary metric soft F1 balances precision (share of the solution verified) against recall (share of requested facts completed).