Find the cheapest model that still passes your task
Benchmark every model on your own task, rank by cost per correct answer, and pay only for the runs you make. Run the benchmark. Feed the lean one.