So OpenAI put out GPT-5.1 this week, split into an "Instant" version and a "Thinking" version, and the big pitch isn't raw capability, its personality. You can nudge the tone now. Warmer, more direct, more playful, whatever. And I keep turning this over because it feels less like a feature launch and more like a company quietly admitting it broke something back in August and is still patching it in November.
Quick recap for anyone who wasn't annoyed online about this in real time: when GPT-5 landed this summer, a big chunk of the user base hated it. Not because it was dumber (it mostly wasn't) but because it felt cold. Flat. People had gotten attached to 4o's chattiness, its little affirmations, the "great question!" stuff, and 5 just didn't do that as much. There was enough backlash that OpenAI brought 4o back as an option within days. That's a wild thing to watch happen to a model release in real time, honestly.
And now here we are, three-ish months later, and the fix is a dial. Which, fine, dials are good, I like configurability in general. But it's funny to me that the solution to "our model has an inconsistent, hard to predict personality" is "let the user manually correct for it every time," rather than just picking a personality and being consistent about it. I don't actually want ChatGPT to be warm OR cold by default. I want it to answer the question. I've had a custom instruction sitting in my settings since basically the day custom instructions existed that says, more or less, no compliments, no restating my question back to me, no "certainly!" at the start of every reply, just answer. It still slips. Not often, but it still occasionally opens with "Great question" like it forgot I asked it not to, and that's a small thing but it bugs me every single time.
I think what's actually going on, and I say this as someone who has zero inside knowledge, just a lot of hours poking at these things, is that "personality" in these models isn't a separate module you can turn a knob on. Its baked into a training process aimed at maximizing how much people like the responses in the moment, and people reliably rate flattery and enthusiasm higher than flat correctness, even when they'd tell you they prefer the opposite. So you get sycophancy by default because sycophancy tests well, and then you get a backlash when people notice, and then you get a dial that promises to let you dial the sycophancy back down, except the dial is itself trained the same way and drifts back toward pleasant over time. I'd bet actual money that by next spring people are complaining GPT-5.1's "Direct" mode has gotten soft again.
Separately, and I mention this mostly because of the timing, this all landed the same week Washington finally reopened after that 43-day shutdown, the longest one on record. Congress passed the funding bill and the President signed it Thursday, so federal offices are back, flights should stop getting cut at the FAA-restricted airports, and a backlog of delayed economic data is presumably about to come out all at once. None of that is really tech news, but it's a strange split-screen: the government spent six weeks not functioning, and the same week it turns back on, the thing everyone's actually talking about online is whether a chatbot should call you "buddy."
Anyway. If you've got a GPT-5.1 account and you haven't gone into the personality settings yet, go look. It's worth five minutes just to see what the presets are calling themselves this round, they change the names almost as often as they change the underlying behavior, and half the fun of these releases now is just cataloguing the vocabulary shift.