Zhang et al. (2026) built ParseBench from roughly 2,000 human-verified enterprise pages spanning insurance, finance, and government. The original five-dimension study found no method strong on every axis; LlamaParse Agentic led that full comparison at 84.9%. Vendors may publish a subset: Cohere's August 2026 Parse 5 announcement reports a three-dimension average (tables, content faithfulness, semantic formatting) under updated August 2026 formatting rules, and excludes charts and layout because those fall outside Parse's product scope.