Wasting two tokens on start/end reasoning seems expensive to me (a priori) I am curious what that would yield though - in some ways that would be the most fun

Wasting two tokens on start/end reasoning seems expensive to me (a priori) I am curious what that would yield though - in some ways that would be the most fun to analyze (when does it think a lot??) I would also be curious to see at what point you see diminishing returns from reasoning tokens (eg a 1:10 ratio? More?)