It’s not cheaper to run Claude in your own GPUs rather than the $200/mo for certain workloads. For a large portion of what I work on, the bottleneck is my time, not tokens. You certainly could throw more tokens at it but if you need it to work a certain way for certain reasons, and your plan/goals are beyond the scope of what the top-capability models can do, then throwing them at the problem just bogs you down in extra cruft or reviews/iteration that you could more effectively do being the primary driver of the work.