What does the comparison actually measure?

“Cost per successful task first, then latency, reliability and explainability. Both sides run the same tasks with the same models and tools under equivalent quality and policy requirements, which is what isolates the value of the execution layer.”