Measurements, traps and the occasional argument. No roundups.
I measured thirty days of real coding-agent sessions. Output was 14 percent of the cost. The cached prefix was 84. That changes what is worth optimizing.