C++ can do a lot of things but one thing it can't do is perform as poorly as python.

This is a very "yes but still no" thing in my experience. Typically, I find that if I write "naive" C++ code, where I make no effort to optimise anything, I'll outperform python code that I've spent time optimising by a factor of 10-30 (given that the code is reasonably complex, this obviously isn't true for a simple matrix-multiplication where you can use numpy). If I spend some time on optimisation, I'll typically be outperforming python by a factor of 50+.

In the end, I've found it's mostly about what kind of data structures you're working with, and how you're passing them around. If you're primarily working with arrays of some sort and doing simple math with them, using some numpy and scipy magic can get you speeds that will beat naive C++ code. On the other hand, when you have custom data structures that you want to avoid unnecessarily copying, just rewriting the exact same code in C++ and passing things by reference can give you massive speedups.

When I choose C++ over python, it's not only because of speed. It's also because I want a more explicitly typed language (which is easier to maintain), overloaded functions, and to actually know the memory layout of what I'm working with to some degree.