The model isn't your bottleneck. What you build around it is.

Most teams answer a disappointing result by reaching for a pricier model, and burn budget without closing the actual gap.

I hand a frontier model a small app to build from scratch, no harness, and let it one-shot it. Then the same task goes to a cheaper model, wrapped in a harness (guides and sensors) I build live, layer by layer, until it matches the frontier run. The sharpest layer: a second agent, with no memory of how either version was built, that reviews both outputs and judges whether they hold up.

This is for developers using AI coding agents who want to move from prompting to designing the system around the model. You'll leave with an ordered recipe: which layer to add first, what it costs, and what it buys you. That's how you decide where your team's effort belongs, priciest model or not.