I've found Claude to be the least sycophantic OOTB. But I've also been running custom system prompts for a while now and have lost touch with what the default behavior looks like. I've had to balance "push back when I'm wrong" and "act like a peer" style instructions which have lead to some pretty annoying one-upmanship style dialogue.

More specifically those instructions alone caused the model to narrow in on tangentially incorrect details while completely ignoring the thrust of my argument. I get where this could cause issues with writing code, but I do a lot more than write code with CC and the conversational voice is important to me.