Of course it depends what you're trying to do. I'm pretty sure that numpy/scipy have a lot more overhead than Matlab for everything that _isn't_ just a blas/lapack call, but I don't have the numbers to back it up.

I do know that I tried rewriting one of my C++ codes using scipy/numpy and it was about a hundred times slower for that particular task (which did happen to involve very large matrices, approaching the limits of memory).

From experience that kind of slowdown for pure numeric algorithms happens when your numpy/scipy code is structured in a way that it makes extra unnecessary copies.