I ran some test with gpt- 5. 6 sol in max reasoning on both codex cli and my own agent harness: https://github. com/Tura-AI/tura I tested only 1 task and the toekn efficency difference is not as great as in high mode: Tura used up to 83.
Source: [Hacker News](https://news.ycombinator.com/item?id=48947713)