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 83.1% fewer turns on the rewrite benchmark and improved the harness results by 7 percent point. In ultra the dirrence is around 50% and incressed harness result only by one. Full report: https://turaai.net/blog#is-gpt-5-6-sol-max-worth-it

However in the bug fixing tasks based on the pulbished report from DeepSWE and others, the performance incress is less impressive compare with the incess in costs. here is the summarry:

Use Max for building new project, and high for debuging or add defined feature.