Less tokens. These models already overthink like crazy especially for complex tasks.