If it costs the same compute to run it then there is no point running worse models

"Worse" model is largely subjective. Often, task specific.

For me, I find model upgrades frustrating as they often break subtle things about my workflows while not clearly offering an improvement. It takes time to learn the nuances of each model and tweak your prompts to get the best outputs.

For example, Sonnet 4 is now my daily driver for Cursor - but it took me nearly a month to tweak my approaches I was using for 3.5 and 3.7.

Unless you invest to self host models that is going to be the case forever, expecting yearly breaking changes is realistic.

That's assuming all else holds on the model which isn't always clear.