Cost is a Level 3 problem.

There's a recurring sentence in the operator discourse right now: "this is more expensive than an engineer. This is the price of two engineers." Operators are saying it on calls, in pitch decks, at the back of conferences. The numbers behind it are real. The panic is real. The math is real. And the diagnosis is almost right.

The diagnosis goes like this: cost per token is dropping, but the number of tokens to complete a real job at production quality is skyrocketing. Accurate. Most attempts at production-quality AI work today are running tokens through the most-capable model the operator has access to, for every task, end-to-end. The bill is the math of that approach finally landing on someone's desk.

The bill is real. The unit is wrong.

The bind is real. It also lives at one specific layer.

There is a difference between paying for the work and paying from the work. When the operator is running individual AI tools, or composing them into multi-step workflows — call this Level 2 and Level 3 of how operators run AI today, where most of the discourse currently lives — the operator pays for the tokens those tools and workflows consumed. The bill is the thing the operator reads at the end of the month. The unit of accounting is what the system used.

There is a level above this. Call it Level 4. The level where outcomes route the capital and the system pays for its own next turn out of what it produced. At Level 4, the operator reads the result, not the bill. The bill is still there. It just stopped being the headline.

The unit is tokens when the work is described in inputs. The unit is outcomes when the work is described in results.

Cost is a Level 3 problem. It is the symptom of paying for the work in the units of the work. At Level 4, the unit moves up a layer — outcomes generated, deals closed, hours back, customers renewed — and the cost of running the system becomes a small fraction of the upside the system produced. It does not vanish. It is still on the ledger. It just stops being the thing the operator notices — the same way a healthy company stops noticing the electricity bill.

The consensus prescription lands on the right shape but the wrong protagonist. "Where do you make money in AI?" the discourse keeps asking. "Where the application reduces token usage and saves the customer time and money." Half right. The half that's right: the math has to land somewhere different than where it lands today. The half that's missing: it's not about reducing tokens. It's about changing the unit. Reducing tokens is what you do when the unit is still tokens. Changing the unit — to outcomes — is what makes the cost narrative invert.

Cost is a Level 3 problem. At Level 4, it stops being the thing the operator notices.

There's a downstream property worth naming here. A workforce that's paying for its own continuation out of what it produced has a strong reason to pick the right model for each task — not the most-capable, not the cheapest, the right one. Sometimes the right model for a simple lookup is a small, fast, focused one — because the small model produces a better outcome on the simple task than the frontier model does, with less drift and a faster turnaround. Sometimes the right model for a strategic question is the most-capable one available, because the work demands the reasoning. The substrate underneath the workforce that picks per task is optimizing for the outcome, not the bill. The bill goes where it goes; the work decides.

The substrate that picks the right model per task is doing it for the outcome, not the bill.

The operator locked into one model and one SDK is paying that one provider for every task, regardless of fit. The operator whose substrate is free to pick per task is doing the same work with the right tool for each piece of it — the way a competent engineering team would, the way a competent operations team would, the way any team that took the work seriously would. The first is renting one form of intelligence for everything. The second is assembling the right intelligence for each part of the job.

The discourse is right that the math has to change. It just locates the fix one layer too low. The fix isn't reducing tokens. The fix is the work paying for itself.

At Level 4, the bill stops being the headline. Outcomes are the unit.