I think two things happened:

1. The sheer number of tokens that a coding agent can use flipped the math upside down on this equation. If you use the most expensive model for everything those costs quickly become untenable, even for software companies.

2. We realized many of the coding problems we're solving aren't incredibly difficult.