They can be. I've done some drift detection work against local models and for the most part, they do ok. I think there's always room for an augmented approach where local models can handle programmatic parsing and structure and large models can handle actual coding routines. I try to use witness coding using local+api where possible to see if there's capabilities that can be caught one side or the other.