I agree that adjusting the reasoning effort to retain cache is a huge thing! But even doing that automatically is currently a challenge for people to figure out and do well, and costs mental energy when perhaps it doesn't need to. For example, there is GPT-5.6-Sol low, med, high, xhigh, max, and lots of "rules of thumb" that people develop on which one to use when.

caching is per model, it does not transfer between them