The Cheap Model Became the Default and Nobody Noticed the Downgrade

I caught myself routing every task to the cheapest model this month and only then wondered if the work had suffered. It had not, and that tells you more about the old pricing than the new models. Most of what I ask a model to do is summarize, draft, or classify, and those jobs did not need the flagship.

The price cut that did it was GPT-6.1 Terra at roughly half the cost of its predecessor. Once the agent step costs a rounding error, the instinct flips from which model is best to which model is cheapest that is good enough, because good enough at a tenth of the cost wins every spreadsheet.

There is a real risk in this habit, and I know it. The cheap model is worse at the hard thing, the multi-step reasoning or the careful code, and the failure hides because most tasks are easy. The discipline is to keep a list of the few tasks where the upgrade earns its price, and never let the cheap default quietly handle one of them.

The shift is healthy. We spent two years assuming the most expensive model was the right default. The market just told us most of our work was never that hard, and the money we save is the proof.