I Turned the Reasoning Dial to 'High' on 4 Models. It Fixed One Thing and Billed Me for Everything.

Same 60 questions, same prompt, same grader. Only reasoning_effort changes. 1,800 graded calls later: one real accuracy gain, many bigger bills, and a model that counted two tools as 10,016.