I Stopped Chasing the Leaderboard and Picked by Price
Benchmark one-upmanship stopped changing my output. The model that's good enough and cheap enough won, every Tuesday.
I used to switch models whenever a new one topped a benchmark, rewriting my prompts each time. The outputs barely moved. The frontier is a staircase where each step is noticeable but not decisive for most of what I build, and the churn cost me more than the gain.
So I picked by price and fit. A $2 model that does the task at 95% of the frontier is the default; the frontier is the exception I reach for on the hard 5%. My per-token bill dropped and my prompts stopped dissolving into version notes.
The discipline is revisiting quarterly, not weekly. The gap between tiers compresses (see this month's rate cards), so the 'good enough' model keeps getting better while I'm not looking. Chasing the top means re-paying the switch cost for a margin I can't feel in the product.
Benchmarks are marketing with numbers. The model you should use is the one whose bill you stop noticing and whose output your users don't complain about. I found mine by price, and the leaderboard stopped being my homepage.