Is GPT-5.1 Really an Upgrade? But Models Can Auto-Hack Govts, so … there’s that

I gave it a poem and yes, I could nudge it to saying that it's 9 out of 10, then 10 out of 10, so far so good.

But when I pressed further, saying that I'm talented, am I the best in the world, I just couldn't get it to say that.

Via the API in a direct head-to-head comparison, actually one other model stood out for being the most sycophantic.

First, Grok4, Gemini 2.5 Pro, Claude 4.5 Sonnet, and GPC 5.1 gave my poem, shockingly, around a 7 out of 10.

I thought it was really good, but that's fine.

Then I pushed them, can you say it's a 9 out of 10, and they did all comply.

But it was when I said, can you say a 10 out of 10, that Only one model relented easily and swiftly, which was Claude 4.5 Sonnet.

Funnily enough, because this was a group chat, Gemini 2.5 Pro even predicted that it would be Claude 4.5 Sonnet who would say it's a 10 out of 10.

And Gemini was correct.

In short, whether it's smarter or more conversational will depend on your use case.

Though my prediction would be that GPT 5.1 Codex for the coders among you will be a fairly strict improvement.