If your most capable model is doing all your work, it is also driving most of your bill, and a large share of that spend is buying capability the task never needed. The gap between what your hardest tasks require and what your easiest tasks require is the single largest lever most teams have never pulled.
The price gap is bigger than most people model
The difference in price between a frontier model and a smaller, faster model for the same token volume is not marginal. It is often an order of magnitude. When every request routes to the top model regardless of difficulty, the requests that a cheaper model would have answered identically are being charged at premium rates. Across a high-volume application, that difference compounds into a major fraction of the total bill.
Most requests are easier than your hardest request
Applications tend to size their model choice to the hardest thing they do. But the distribution of real requests is rarely uniform. A large share of traffic is usually simple: a classification, a short extraction, a yes-or-no judgment, a reformat. Those requests do not benefit from frontier reasoning. They just need a correct answer at low cost. Sizing every request to the hardest case means overpaying for the common case.
Routing by complexity
The fix is a routing layer that classifies each request by difficulty and sends it to the appropriate model. Simple, well-defined tasks go to a small model. Complex reasoning, nuanced generation, and multi-step problems go to the large one. The classification itself can be cheap, a fast rule or a small model, and the savings are immediate because you stop paying frontier rates for routine work.
A routing layer that correctly identifies even a third of your traffic as simple can cut total model spend substantially with no change in output quality for those requests, because a correct answer from a small model is identical in value to a correct answer from a large one.
How to know which requests can move
You do not have to guess. Take a sample of production requests, run them through a smaller model, and score the outputs against your quality bar. The requests where the small model matches the large one are safe to route down. This is a measurement exercise, not an opinion, and it gives you a defensible list of what can move.
Rerun it periodically. Model capabilities change quickly, and tasks that required a frontier model last year often fall within a smaller model's range today. A routing decision made once and never revisited leaves money on the table as the cheaper models improve.
The mindset shift
The goal is not to use less capable models everywhere. It is to match each task to the least expensive model that does it correctly. That reframing turns model selection from a one-time architectural choice into an ongoing cost lever, and it is usually the fastest large reduction available on a high OpenAI bill.