Most writing about agent cost is estimation. We run a control plane that meters and prices every call, so these posts are built on our own metered data instead: 285,814 governed tool calls across 96 workspaces, where the orchestration loop was 97% of total spend in the July 2026 corpus and one model drove 77% of the bill from 30% of the calls. The pattern that falls out — the loop tax, the leaf work, the one step that should be 90% of the bill and the runaway session that shouldn't — is the difference between cutting an agent bill and just staring at it.
All the headline numbers, with method and provenance, live on the data page. To price your own coding agent, start at ACP for coding agents.