Sorry, didn't mean they're using pi. They're itemizing ways to get better results but not showing pi, while other testing shows most models (including Opus) achieve better from within pi.