The 'feel' of a single person is pretty meaningless, but when many users form a consensus over time after a model is released, it feels a lot more informative than a simple benchmark because it can shift over time as people individually discover the strong and weak points of what they're using and get better at it.