Opus 4.6 costs 5x what Haiku 4.5 costs per token. Anthropic just gave every Claude user a button to switch between them mid-conversation. Most conversations follow the same arc: you open with a hard reasoning problem that needs the strongest model, then spend the next 10 messages on implementation, follow-ups, and debugging that Haiku handles fine. Until now, every message burned Opus-tier compute regardless. This is what cloud infrastructure learned a decade ago. AWS doesn't make you run…