Overall, I’m still very happy with gpt-5.6-sol. My experiment to lower effort level to “medium” for most of my projects is going well.
Doshmodo, hlsync, discovery engine and pluckso are all using the lowered level now. Million minds AI, on the other hand, I raised to “xhigh” and it really made a difference there. It got complex enough that lower levels are routinely missing edge cases or deeper issues that requires multiple rounds to catch and handle.