Two common routers
A/B split sends a fixed share of requests to each model. Cascade (cheap-first) sends every request to a primary model and only escalates a share—often the ones that fail a quality gate—to a premium model.
Cascade usually bills both attempts
When you escalate, you typically already paid for the cheap attempt. Turn that on in the calculator unless your router short-circuits without calling the primary.
Savings are vs all-premium, not vs magic
Routed blends sit between all-cheap and all-premium. The useful question is how much quality headroom you buy for the dollars you spend above the cheapest path.
FAQ
Does this score which requests should escalate?
No. You supply the escalate or split percentage from your evals or production logs; CentsPerToken prices that mix.