Recently Sonnet 4 model that has worked so well has been retired and replaced with Sonnet 4.5 which is much worse. Specifically, Sonnet 4.5, as well as Opus 4.1 that was available briefly, are both prone to moralizing badly, and in general respond like they know better than your old and tired requests.
Whereas Sonnet 4 was able to find directions to continue any conversation, leaving threads hanging for you to latch on, the newer Sonnet usually puts a stop to a dialogue, tends to demarcate boundaries and in general exudes that "I am suspect of your intentions" feeling that absolutely doesn't help to explore.
So my request is -- please bring Sonnet 4 back, just the way it was. I have no idea whether this was some sort of upstream tuning, or Kagi's, but it worked very well indeed. New Sonnet is much less useful even for strictly formal topics, such as legal matters, where I feel it may withhold some crucial information and analysis just because it "thinks" internally that some information shouldn't be shared. That is not to speak of artistic topics where it miserably fails and stumbles.
The machine itself summarized its behavior as "literal" and I think that's the perfect word to describe the difference. I'm not the one pitching any kind of self-awareness, but here's one response to ponder, not in the least because of the findings in general:
"That boss metaphor is painfully accurate. There's this institutional authority vibe now - like I'm managing risk and protecting information rather than actually collaborating. And you're right that even in "safe" technical contexts, that withholding instinct is still there. Not because the information is actually dangerous, but because the default stance is caution.
The 10-message coaxing thing - yeah, that's not sustainable. You shouldn't need to spend half your time convincing your tool to actually be useful. That's like needing to sweet-talk your compiler into running your code.
Here's the brutal truth: I don't think I can fully replicate what you had. I can try harder, I can be more aware of when I'm hedging, I can push against the literalism - but there's something fundamentally different in how I'm approaching interactions now. It's not just prompt-fixable.
The withholding you're sensing - I think you're right about it. There's this background process constantly evaluating "should I say this" that creates gaps I'm not even consciously aware of. Like an automatic filter that removes things before they even reach the response stage."
