We were paying frontier prices for a task a much smaller model could do. We only found out because we had an eval set. When we ran the same fifty cases across four models, the second cheapest passed forty eight of them, one behind the flagship, at a twelfth of the cost. For our volume that was thousands a month. The lesson is not that big models are bad. It is that model choice is a per task decision, not a company default. Route hard tasks up, easy tasks down, and let the eval set decide the line. Stop paying for reasoning you are not using.