It's a matter of abstraction 'stability'. There's a cost for each abstraction, so modeling without fully understanding what you try to model or something that changes wrt the abstraction model will yield incomplete abstractions, be hard to read, etc.
Many things are being remodeled over an over (incomplete, with bugs, etc.) when in fact they don't change much. These are good candidates for good abstractions. I don't think you'd want to do 3d graphics without a system that understand the algebraic aspects (e.g. orthogonality) behind rotations, translations, etc. Categories generalize this all the way so optimizations would be possible at a whole different level (makes me think about what Tensorflow enables).
I also think you only need one category (just like you only need one group / semigroup concept) for a model and it could be language agnostic.