Token usage also needs to be factored in specifically when thinking is enabled, these newer models find more difficult problems easier and use less tokens to solve.