It would be interesting to give modern students past papers from 2000 1990 1980 down to 1950, well, 51 I suppose, but even more interesting to do it in reverse, though impossible of course! The changes in the papers are well known: the topics are more elementary and predictable, you don't need to structure your proof yourself, there's minimal manual computation, but you do need to decode from text much more, and translate repeatedly into and out of concrete examples, which is more accessible to most people and a drag on purely abstract thinkers.
It clearly makes sense to optimise for "messy modelling" type skills because those are the types of maths we need really widely in modern life. We need humans who can identify what the math means, and how to apply it to a messy real-world constraint. Applied mathematics like statistics and software engineering is what we're doing most of. We don't really need to spend a lot of time hand-computing nowadays.