There's an escape valve: I've had Claude stand-up AI-enabled features in my app, so we're much less dependent on Claude itself. The app uses cheap API calls to check code quality and run other "lessons learned" sweeps. Some of these become regular code in the end as well.

Same. Claude "loves" writing tools to make checks deterministic so much that I actually have to pull it back sometimes and point out that some things need to keep using LLMs. But even then, I have it "farm out" a lot of things to scripts using dirt-cheap models for things I don't need a Sonnet or higher level model for as well. The proportion that runs on the SOTA models keeps dropping.