The jump from 1M to 100M tokens daily is wild. I had a similar experience—my automation agent went from $5/month to hitting weekly API limits when I gave it too much autonomy. The fix wasn't better prompting or using Opus everywhere. It was realizing most tasks don't need the expensive model. Now I route: Haiku for lookups/email, Sonnet for content, Opus only when synthesis matters. Token costs dropped 70%, output quality stayed the same. Turns out the model selection layer is the real infrastructure challenge. https://thoughts.jock.pl/p/claude-model-optimization-opus-haiku-ai-agent-costs-2026
The jump from 1M to 100M tokens daily is wild. I had a similar experience—my automation agent went from $5/month to hitting weekly API limits when I gave it too much autonomy. The fix wasn't better prompting or using Opus everywhere. It was realizing most tasks don't need the expensive model. Now I route: Haiku for lookups/email, Sonnet for content, Opus only when synthesis matters. Token costs dropped 70%, output quality stayed the same. Turns out the model selection layer is the real infrastructure challenge. https://thoughts.jock.pl/p/claude-model-optimization-opus-haiku-ai-agent-costs-2026