The cheapest model beat the flagship on my actual work, and I kept the receipt
A month of side-by-side tasks showed the priciest model is not the right one for most of what I do.
I stopped assuming the most expensive model was the best one. For a month I ran the same tasks on the flagship and the cheap tier and tracked which answer I actually shipped.
The flagship won on the hard, ambiguous jobs: a messy legal summary, a research synthesis, anything where judgment mattered. There, the extra cost bought fewer mistakes.
The cheap model won on the boring eighty percent: rewriting a subject line, sorting a list, drafting a status update. On those, the flagship was slower and no more right. I paid a premium for nothing.
The lesson is not cheap is better. It is that most of my work is the boring eighty percent, and a small model eats it for a fraction of the price. Save the flagship for the twenty percent that bites back.
I now route by task, not by habit. A ten-cent job does not need a two-dollar model, and a two-dollar problem does not deserve a ten-cent guess.
If you only change one thing: stop defaulting to the flagship. Match the model to the mess, and the bill drops without the quality.