Cheaper AI Models for the Grunt Work, Expensive Ones for the Brain

The default habit is picking the smartest model and pointing it at everything. It feels safe. It is also the most expensive way to build.

My read is that most tasks in a real workflow are not hard. They are repetitive. Writing a Supabase query, wiring up a component, adjusting a layout. That work does not need the most capable model on the market.

One routing config in the Claude community measured a 73% cost reduction from a single change. That number will vary by workload, but the direction is obvious once you look at your own task list.

Think of it as one manager and two specialists. The coordinator holds the plan. The backend agent only touches Supabase. The frontend agent only touches the visual side.

The coordinator decides what each one does and in what order. It never writes the code itself. It delegates.

That separation is the whole trick. Once the hard thinking is isolated into the coordinator, everything downstream is just execution. And execution runs fine on a cheaper model.

The split I use is Opus for the coordinator and Sonnet for the two workers. Opus is better at holding context and breaking a messy goal into clean steps. Sonnet is more than capable of carrying out a step that has already been defined.

The per-call gap looks small. Around a thirty-cent difference between the tiers in some measurements. But you are not making one call. You are making hundreds across a session, and most of them are worker calls.

That is where the money leaks. Not in the one big planning call, but in the fifty small ones you kept sending to the premium model out of habit.

Routing is not free of risk. If you push a genuinely complex task to a cheaper model, you get worse output and you end up re-running it, which costs more than if you had just used the smart model once.

So the coordinator has to be honest about what is actually hard. That judgment is the part worth paying for.

I keep seeing people set this up and then route everything to the cheap tier to save more. That defeats the point. The goal is not the cheapest model. It is the right model for each specific job, decided by something that understands the job.

Before you build any of this, look at what you asked your model to do last week. Sort it into planning versus doing.

Most people find the doing pile is much larger. That pile is your savings.

Then set one rule: coordination and architecture stay on Opus, defined execution moves to Sonnet. You do not need a fancy semantic router to begin. A clear split between planning and doing gets you most of the way, and you can refine from there.