It's definitely O(N log(N)). It COULD also be O(N), if you could show that the sequence of calls, as a total, grows slow enough (because while SOME calls are O(log N), enough others are fast enough, say O(1), to bring the total down).
Remember: O(f) means the algorithm is no SLOWER than f, but it can be faster (even if just in certain cases).