I agree. I was completely sold on Claude models for a year. 4.6 vs OpenAI codex in same period? It was night and day. Opus I could talk to about api design, tradeoffs, etc. codex was mechanical, used “load bearing” constantly, and unsettling brief.

Now it’s flipped. Sol emits thoughts as it works, which help as I’m scrolling through and see it’s made a bad assumption. It can be directed but still push back. Opus? It’s seems to inherit the unsettling silence of Fable and waits till the end to give you its authoritative “here’s how it is. I even end up having 4.6 “translate” what it says back to English. I hate having to instruct an llm to “talk to me”.

Yes there’s ways of getting it to talk more plainly, “don’t overwhelm me”, not be as nit picky and anxious “we are bold and fearless”. But didn’t have to do that before.