Does it depend on the model? I would have expected the bigger ones to be better with common sense and not fixating on irrelevant details. But I have only used them with quite small codebases so far. (Which have basically no internals to exercise!)