Price per M/tokens is also a fuzzy metric when newer models reason longer, and then burn more tokens while doing so.