The frontier model is the right default for a new workflow. It is the strongest generalist you can buy, and while a task is still being discovered — while its shape, tools and failure modes are unknown — paying for that generality is correct. The question is what to do when the task stops being new.
Agent traffic repeats. The same task families recur thousands of times a month, and frontier pricing does not care that the thousandth run looks like the first. Inside one model family the price spread is roughly 5× — which only pays if you can prove the cheaper model does the specific work well enough. The proof is the hard part, and it is the part most teams skip.
Three preconditions
A move down the ladder is justified when three things are true at once. The task repeats — enough volume that the saving matters and enough examples that a skill can be learned from them. The eval is stable — a held-out set and a definition of good that does not shift under you. And the economics bind — the cost or latency of the frontier route is actually limiting the product, not merely annoying someone.
Note what is absent from that list: the cheaper model being good at everything. It does not need to be. A weighted A/B across all traffic asks "is the cheap model good enough for everything?" — and the answer is usually no. The workload question is different: which specific slice of the work is it good enough for, once it carries the skill that slice requires?
The arithmetic got harder, not easier
A naive model switch can cost more than it saves. Cache reads price at a fraction of fresh input, and switching models abandons the warm prefix — so the arithmetic has to be done in the request path, per turn, not on a spreadsheet afterwards. This is one reason routing decisions belong in a gateway that can see the cache state, not in application code.
The gate is the decision
Operant's rule is that the cheaper route takes traffic only after it matches or beats the frontier baseline on held-out work, and a person ratifies. Until then, it runs in shadow — visible and priced, serving nothing. After promotion, the frontier model remains the failover and demote is never gated. A move down that cannot be reversed in one click was never a safe move at all.