The experiment provides concrete token-count measurements showing that schema design and output pruning — not model choice — are the dominant levers for reducing MCP call costs, with output pruning alone responsible for 35–40% of total token overhead.