You start every new workflow on the top model, and that's the right call while you're still learning the task.
But agent work repeats. The same kinds of task come back thousands of times a month. The top price doesn't drop for the thousandth run.
Operant moves a task to a cheaper model only when the move has been proven on your own work. Inside one model family, the price gap is roughly five times. But that only pays if the cheaper model does your specific work well enough. The proof is the part most teams skip.
Three things have to be true
A move down is worth it when all three of these hold at once:
| Condition | What it means |
|---|---|
| The task repeats | Enough volume that the saving matters, and enough examples to learn from |
| The test is stable | A set of unseen chats, and a definition of good that doesn't shift under you |
| Cost binds | The cost or speed of the top model actually limits your product |
It doesn't need to be good at everything
Notice what's missing from that list: the cheaper model being good at everything. A 50/50 split of all your traffic asks whether it's good enough for everything. The answer is usually no. The better question is which slice of the work it can do, once it has the instructions that slice needs.
The math happens on every turn
A careless model switch can cost more than it saves. Cache reads cost a fraction of fresh input, and switching models throws the cache away. So the math has to happen on every turn, in the path of the request, not in a spreadsheet afterwards. That's one reason this decision belongs in a layer that can see the cache, not in your app code.
The test is the decision
The cheaper model takes traffic only after it matches or beats the top model on chats it never saw. Then a person has to approve the move. Until then, it runs without serving anything, visible and priced. After the move, the top model stays as the backup, and moving back needs no test.
So a repeated task moves only after the cheaper model matches the top one on your unseen chats. Then a person has to say yes. A move you can't undo in one click was never a safe move.
The top model keeps each task until a cheaper one proves it can do that exact work.
See how that proof works in How we prove a cheaper model can do the job.