In my case, I don't use it for programming (coding), but rather build an LLM Wiki together with M365 MCPs integrated into Github Copilot CLI for work purposes. Since March, I've been applying it in earnest and have been using almost only Opus so far (4.6~4.8)
Today, GPT-5.6 (sol) was finally updated, and after using it for a few hours, the feeling is that it does work carefully and well. When I request to start work every morning, Opus scrapes all the information, presents the next steps, and proceeds by asking me questions, but GPT-5.6 just neatly organizes only the content determined in the instructions and ends. This part has pros and cons. While it seems good that Opus does things well and naturally, since it also sets priorities on its own and keeps moving forward one step at a time, requiring me to make many corrections, GPT-5.6 reduces back-and-forth since I just need to give the next steps, and I also like the feeling that it still does things well appropriately within the content I instructed.
And regarding deliverables, even when I ask Opus to express things cleanly, no matter what engineering presentation slides, the titles have overly dramatic expressions and too much emotional language, so I had to make many corrections. But GPT writes with appropriate expressions that fit the content perfectly, which is nice. This seems to be related to differences in expression preferences depending on the related work.
Anyway, in-house Fable is disabled, so it's either Opus 4.8 or GPT-5.6 Sol, and with that I'm leaning heavily toward 5.6. I suspect it's because Fable's token usage is too wasteful. When running similar tasks, GPT-5.6 and Opus 4.8 feel almost the same in terms of AI credit usage.
