Someday (probably), a model will be trained once that is better than a human at coding, and it will only need trained once.

It can then be used indefinitely for the cost of inference, which is cheap and will continue getting cheaper.