They are nearly identical in speed. The 'anonymous' method has to form a temporary user vector, which takes time proportional to the # of items passed, but this is quite fast. Neither uses any significant memory.
The problem with always using the anonymous method is that you would have no data actually in your model then!