It depends on what your "currency" is: inference cost vs. models getting dumber/slower with a fuller context.