Notes on measuring what automation and AI return, written while building the thing that measures it. Mostly numbers, and the occasional account of getting one wrong.
Per-token pricing has to be a dated table with a version stamp, and a financial input needs a floor under it. Here is what my sync keeps, what it throws away, and the lookup miss that taught me to leave a run unpriced.
Metering what an agent spends is a solved problem and several good tools already do it. Working out what the run was worth is the half that goes wrong, and it goes wrong quietly.
Nine Claude Code sessions cost twelve cents to run. The time-saved figure I got first was double the truth, and the reason is the most common way people overstate what their agents are worth.