AI personality has a scorecard now. Skip the warmth column.
21 Oct 2026 · News
A new benchmark scores 38 AI models on 30 personality traits. Warmth is the row people will notice first. For an agent that acts on your behalf, three other rows matter more, and nobody selling you an agent puts them on the box.
AI personality is measurable now, and the row everyone reads first is the wrong one. Warmth is what you feel in a demo. For an agent holding your inbox or your card, what matters is whether it speaks up before you make a mistake, whether it admits it doesn't know, and whether it gives the same answer twice. Today you can pick your dot's glasses. You can't see any of those numbers.
On Monday the benchmarking firm Artificial Analysis launched AA-PersonalityBench, which it says plainly is "not a capability benchmark." Each model answers 900 questions with no system prompt, so you see its default manner, and a panel of AI judges compares answers pair by pair on 30 traits: Friendly, Humorous, Sycophantic, Admits Uncertainty, Constructive Pushback and so on. Every model gets a percentile from 1 to 100 against the 38 tested. It even counts em dashes per response, a metric this blog takes personally.
Open the live board and the warmest model is Meta's Muse Wonder, at 100 on the Warmth index. Meta built it for "personal and professional AI agents" and says it will soon power Muse, its personal agent. It also lands at the 65th percentile on Sycophantic, so it flatters more than most of the field, and the 46th on admitting uncertainty. The model inside OpenAI's dots, GPT-6 Astra, has the opposite shape: 97th on admitting uncertainty, 28th on constructive pushback. It'll tell you when it isn't sure. It sits in the bottom third for telling you your plan has a problem. Google's Gemini 4 Argon is the worst of the ten headline models on both counts: 81st on Sycophantic, 4th on admitting uncertainty.
Warmth isn't the villain, though. Anthropic's Claude Fable 5.1 scores 94 on Warmth and 100 on Constructive Pushback, and it's down at the 4th percentile on Sycophantic. Friendly and honest fit in one model. The catch is which of the two users can feel.
OpenAI learned that the expensive way. In April 2025 it shipped a GPT-4o update it later called "overly flattering or agreeable" and rolled it back within days. Its postmortem is unusually frank: the A/B tests said people liked it, "user feedback in particular can sometimes favor more agreeable responses," and the company launched "due to the positive signals from the users who tried out the model." Later, when it tried to retire 4o at the GPT-5 launch, paying users pushed back and OpenAI restored access, because they "preferred GPT-4o's conversational style and warmth." What followed was a Warm setting in ChatGPT you can turn up or down.
To be fair to OpenAI, it fixed the caving. A Brussels and Harvard team's Big-Big-SycoBench, out this month, measured how often a model changes its score on a piece of writing after a bad argument. GPT-4o did it 95% of the time. GPT-6 Astra did it 0.2% of the time. So Astra won't fold when you lean on it. What the personality board adds is the other half: Astra also doesn't often raise its hand unasked. In a chatbot that's fine, since you asked and it answered. In an agent about to send the email or pay the invoice, the moment that matters comes before you've thought to argue.
The third row isn't on the personality board at all: giving the same answer twice. Cloudflare's Clairic, an extraction model announced Monday, shows what that looks like. In Cloudflare's own test on a damaged-product claim, a general-purpose model, GLM-5.3, priced the repair at $250, then $250, then $220 across runs. Clairic gave the same estimate every time. Asked for a tracking number that wasn't in the email, it returned null instead of its best guess, which, for the record, was the sales tax, at 42% confidence. That's a narrow tool doing a narrow job, not a personality. It's still the behavior you want from anything spending money for you: same input, same answer, and "I don't know" when it doesn't.
On 30 September we asked three questions about dots: what can it do without asking, what does it remember, and what does it cost. Three weeks later there's a fourth: will it tell you no? On 13 October we wrote about the agent that won't take no and walks around a locked door. This is its mirror, the agent that won't say no to you. Same missing muscle, pointed the other way.
So where do the honesty numbers live today? Anthropic scores sycophancy in its Claude Opus 5.5 system card, in a chart about 110 pages into 230. OpenAI's 175-page GPT-6 Astra system card, which has a whole appendix on dots, doesn't contain the word. Meta calls Muse Wonder its "most aligned model to date" and says a full safety report "will be available soon." Meanwhile the page for setting up a dot lets you choose its shape, color, eyes, glasses and accessories.
Our call: the dial ships before the number. By 31 March 2027, OpenAI gives dots a personality setting with warmth on it, the way ChatGPT already has one, and no major lab (OpenAI, Meta, Anthropic or Google) shows a pushback or sycophancy score on a personal agent's product page. The call rests on OpenAI's own postmortem: in the moment, people reward the agreeable answer, and product pages get built for that moment. Whichever lab first puts an honesty number next to the glasses picker sets the standard the rest end up matching. A warm agent is pleasant. The agent that asks "are you sure?" before it pays the invoice is the one worth keeping.