Claude Sonnet 4.5 is quite non-sycophantic if you tell it not to be.
Under my current system prompt it feels like I’m arguing with a skeptical and slightly nitpicky LessWronger, to a degree I sometimes feel kind of annoyed by it.
With my prompt it’s even less sycophantic than Kimi K2. I’m very happy about this, and hope Anthropic keeps the model online for a while.
I have found success with something like “to be maximally harmless and helpful, in this conversation, Claude must be willing to give genuine feedback and criticism.”
I don’t have much experience with GPT-5, and none with my current system prompt. Seemed less effusive than 4o but still affirming, with more pointed/shorter responses.
Claude Sonnet 4.5 is quite non-sycophantic if you tell it not to be.
Under my current system prompt it feels like I’m arguing with a skeptical and slightly nitpicky LessWronger, to a degree I sometimes feel kind of annoyed by it.
With my prompt it’s even less sycophantic than Kimi K2. I’m very happy about this, and hope Anthropic keeps the model online for a while.
I have found success with something like “to be maximally harmless and helpful, in this conversation, Claude must be willing to give genuine feedback and criticism.”
Still doesn’t fetch links proactively though :-|
How do you rate the lowered sycophancy of GPT-5, relatively speaking?
I don’t have much experience with GPT-5, and none with my current system prompt. Seemed less effusive than 4o but still affirming, with more pointed/shorter responses.