It works much better on regular software development e.g. for complex refactoring where cheaper models would produce a lot of garbage results.