Fuck AI

Multiple things can be true at once.

The scenario where it was incredibly expensive for a task at work was because someone rigged it to do it until all the tests pass, and it got in a loop where it never passed. So they spent a lot of money to get nowhere.

This is the workaround for it not being very good that drives expense. It's not expensive to generate code once, but it frequently fails. There a non zero chance that a retry with stuff about the failure in the context window will work, so the current situation is you are supposed to have the correct state as perfectly modeled with test cases, example implementation, whatever and then rig it to retry until it works. Which is a lot of work, frequently impossible to really do, and frequently still ends in failure so in practice you probably should cap it at 3 or 4 retries.

But this particular thread discussed cost, so other issues were off topic.