This is a hard comparison because pretty much all of these libraries call a BLAS implementation under the hood, so what you might be comparing is actually the cost of stuffing the array with your data in each language (which is obviously task specific).
I think it'd be difficult to argue one way or another: Eigen gives you flexibility one way (custom kernels) and Python gives you another (interactive debugging and easy visualization). Cython/Numba make it even more muddied.
I can definitely takes the cake for an easy to use syntax for doing all kinds of linear algebra. It's only able to do that because the extensive operator overloading that C++ allows. However, don't want that for you for a speed. Using something like MKL directly is always going to be faster than eigen. The trade-off is the API is much more complex.
I think it'd be difficult to argue one way or another: Eigen gives you flexibility one way (custom kernels) and Python gives you another (interactive debugging and easy visualization). Cython/Numba make it even more muddied.