What is model routing?
Model routing definition
Model routing is the practice of sending each task to the cheapest AI model tier that changes nothing about the outcome. AI models come in tiers that differ in depth of reasoning and in price per token, often by an order of magnitude or more between the smallest and the largest. A system that sends everything to the largest tier pays for reasoning it does not use; one that sends everything to the smallest fails on the work that needed depth. Routing decides, task by task, which tier the work goes to.
The decision is made on the kind of task, not case by case at run time. Searching a codebase, matching patterns, reformatting, summarising and classifying go to a fast, cheap tier. Architecture, debugging across many files and anything with a judgement call go to the deepest. Between them sits a middle tier, for domain reasoning and review.
Traide's routing rules
Two standing rules shape Traide's own routing. Routine work, such as exploration and search, runs on a tier that costs about 97% less per token than the default model at today's list prices. And anything touching money or payments never drops below the middle tier, however routine it looks. The saving is applied by convention, never isolated on a bill; the post on how Traide builds with AI gives the figures and what they rest on.
Why it is a convention enforced by tooling
A routing rule that lives in a document is followed on the days people remember it. At Traide the rule lives in the tooling: a checkpoint reads the task before a model is chosen, and asks before it lets financial work run on a cheap tier. The rule holds at the same height on a quiet afternoon and the night before a release. Model routing done this way is one of the AI guardrails, and the evaluation harness is what proves a cheaper tier really changed nothing about the outcome.