30% Drop In o1-Preview Accuracy When Putnam Problems Are Slightly Variated