I thought each new generation typically used more reasoning tokens?
It’s on by default, in a way. You can probably prompt these models with “and don’t reason about it, just give me the answer” and probably get a comparably good response without it using reasoning tokens for many things.