the quality-per-dollar inversion is the key insight. cheap model + no context = expensive mistakes. full-context model doing less = actually cheaper when you account for wasted cycles. memory is compute.