Test results are ranked by actual total cost per thousand queries; token counts come from real answering-call billing rather than estimates, and prices use list pay-as-you-go rates.