I created a test to compare reasoning effort levels with Claude. In 44 of the 52 tasks, the high effort setting used more tokens. An adversarial judge found no meaningful difference between high and medium.