Depends on the problem domain. A good example is 1D numerical perturbation theory; my colleagues struggle to beat mathematica over the head until it gives them roughly what they want. On the other hand, it's much easier using jupyter, numpy, and matplotlib.

>Do you think these aren't going to have errors in them?

Of course they do, but they won't go unfixed for years on end. You can also figure out what's wrong without having to reverse engineer a big binary blob.