I wonder if model router adoption is higher for LLMs performing in-app functions, instead of the author's case of writing software?

One of the ways I can make a custom app/harness for a given task more of a competitive advantage is by striving for model agnosticism - the thing works and so has the valuable part outside of a model.

We've already seen older models being unavailable, and the newer (supposedly better) model doesn't do the work as well for some reason. Or a service is just down for minutes or hours.