I think cost is lagging indicator. (it truely does)

I am currently tracking tool count exposed, schema-token size, tool-call failure rate, duplicate tool calls, retries, wall time, tokens before first useful external observation, and “answered without required tool evidence", and some more.

I strongly recommend single metric: task success with evidence, segmented by task type.