Cascade routing tries Haiku first and escalates to Opus on low confidence. What is the break-even condition where cascading is cheaper than always using Opus?
Answer: Haiku must handle >50% of cases successfully
Cascading pays for two calls (cheap + expensive) on the escalated cases. The math: if Haiku handles >50% of cases, the average cost (50% × Haiku + 50% × (Haiku + Opus)) is less than always using Opus. Below 50%, the doubled-call cost on failures exceeds the savings on successes.