The 15x Cost Trick: Frontier Planner, Cheap Worker

Here is a result worth sitting with. A team of AI agents rebuilt SQLite, one of the world's most widely used databases, working from its 835-page manual, and produced a fresh version in Rust that passed 100% of a held-out test suite. Impressive on its own. But the eye-opener was the cost: the very same job ran up to 15 times more expensive depending only on which mix of AI models was used. Same output, wildly different bill. The cheapest recipe reveals a design pattern that is quietly becoming the backbone of serious AI systems, and it changes how you should think about paying for AI.

Planner and worker, not one model for everything

The instinct most people have is "use the best AI model for the whole job." This research shows that is usually the most wasteful choice. The efficient pattern splits the work by difficulty. A powerful, expensive frontier model plays planner: it does the genuinely hard thinking, breaking the problem down, making the design decisions, resolving the trade-offs. Once it has turned a fuzzy goal into a clear, detailed set of instructions, cheaper worker models carry out the bulk of the execution. As the researchers put it: once a frontier planner has collapsed the ambiguity into explicit instructions, less expensive models simply have to follow them.

Why the savings are so large

The reason this works comes down to where the effort actually goes. In a big task, only a few moments genuinely require frontier intelligence, the initial decomposition, the key design calls, certain trade-offs. Everything after that is following the plan, and the vast majority of the tokens (what AI is billed on) are spent there. So if you run the whole thing on the priciest model, you are paying premium rates for a lot of work that never needed premium intelligence.

It is the same instinct we described in judging AI by intelligence per dollar, taken to the system level: don't buy frontier intelligence for work that doesn't need it.

Why this is where AI expertise pays off

This is not a trick you get for free, and that is precisely why it matters for choosing who builds your AI. Driving these savings requires two things: knowing the domain well enough to tell which steps are hard, and the ability to work fluently across multiple model tiers. Whoever can do that, in coding, finance, legal, healthcare, life sciences, and beyond, can profitably take on larger workloads that would otherwise be too expensive for a customer to justify. It is a strong sign of where the real value in applied AI now sits: not in access to a single clever model, but in orchestrating many of them well.

Where this leaves you

You do not need to hand-route between models yourself, but you should know the pattern exists, because it quietly determines what you pay. When you buy or build AI, treat "how do you manage cost across model tiers?" as a real question. A partner who runs the most expensive model on every step is handing you a bill 15 times larger than it needs to be; one who plans with a frontier model and executes with cheaper ones is doing it right. Multi-model systems are becoming the default design for serious AI, and the businesses, and vendors, who master them will do more, for less, than everyone still paying premium prices for routine work.

Frequently Asked Questions