Caveman hurt model performance. If you need a dumber model with less token output, just use sonnet-4-6 or other non-reasoning model.

Does it? I'm not sure I'd necessarily notice but I haven't found it noticeably worse.