For the kind of ML work I do Cursor's "Auto" was too unreliable. It would frequently try to solve relatively difficult tasks using Composer, when it should have been routing to a true frontier model. Annoyingly, doesn't tell you what model it's decided your task deserves, so I often wasted time working through a problem with it, only to realise I was on a dumb model. Then I revert all the work, switch to an expensive frontier model and pay expensive API prices to actually solve the task.

All that is to say, model selection is the main control we have over quality. Giving it up in the name of cost saving will bite you in the long run, especially when Claude Code still has such good plan pricing.