Lived this exact loop. Switched my AI agent to Opus for everything because "it's the best model." Costs hit $200/week. Moved to Haiku for simple tasks - costs dropped 59% but quality tanked on complex reasoning. The whack-a-mole.
What finally worked: a tiered approach. Haiku for log parsing and lookups, Sonnet for most work, Opus only for synthesis. Not one model to rule them all, but the right model for each job.
Lived this exact loop. Switched my AI agent to Opus for everything because "it's the best model." Costs hit $200/week. Moved to Haiku for simple tasks - costs dropped 59% but quality tanked on complex reasoning. The whack-a-mole.
What finally worked: a tiered approach. Haiku for log parsing and lookups, Sonnet for most work, Opus only for synthesis. Not one model to rule them all, but the right model for each job.
Documented the math and what broke along the way: https://thoughts.jock.pl/p/claude-model-optimization-opus-haiku-ai-agent-costs-2026
Your baseline point is spot on - I wish I'd measured before I started swapping.
We built jetty for exactly this reason. I personally witnessed 7 figures go up in smoke for the reason you just described.
Nice! You must have done better job than I did with Wiz :D