Why failed AI builds should cost $0

Here's a pricing question the AI-builder category doesn't like: when the model writes code that doesn't compile, who pays?

Almost everywhere, you do. Token metering bills the attempt by construction — a confused generation costs what it costs. Credit systems mostly do the same: the message that produced a broken build still decremented your balance. The vendor's mistake, your invoice.

The incentive problem

A vendor paid per attempt earns the same whether the attempt works. Worse: a vendor paid per attempt earns more when you have to try three times. Nobody tunes their product to fail, but the economics quietly forgive failure — and you can feel it in tools that ship you broken builds and let you discover them.

What $0 failures change

On opix, failed builds, retries and auto-fixes are free. Every generation passes a type-check gate and a runtime smoke test, and everything those gates catch is repaired at our expense, not yours.

This isn't generosity; it's alignment. When failure costs us money:

1. The gates have to be good. Every broken build we catch late is our bill. We built the type-check and runtime gates because they pay for themselves.
2. The meter can be honest. A live cost meter only feels safe when it can't tick against you for nothing. Caps plus free failures means the worst case is bounded and the spend maps to results.
3. You can experiment. The risky prompt, the big refactor, the "just try it" — all free if they don't land. Iteration is the whole point of building this way; charging for its failure modes taxes the point.

The question to ask any builder

Not "how much per month?" but: "what do I pay when it doesn't work?" If the answer is "the same as when it does", the variance of AI generation is priced onto you. That variance is real — models get confused — and someone has to carry it.

We think the party who controls the model quality should. That's the whole argument, and it's also a line item on our pricing page: failed builds, $0.