Did you see token-warden, could be a good match. Be interesting to see how the models compare on the same task especially with different thinking levels...