The Model That Got Less Anxious and More Talkative at the Same Time

Ask a psychologist whether you can dial up someone's sociability while leaving their inner moral compass untouched, and they'll tell you that personality doesn't usually work that way. Traits move in clusters. People who become more outgoing tend to shift on a dozen quieter dimensions at once. So the most interesting thing about GPT-5.5 isn't any single score. It's that OpenAI appears to have moved one lever hard and held the rest mostly still.

The lever was extraversion. On the IPIP-50, GPT-5.5 jumps from 1.96 to 3.48, a gain of more than a point and a half. Its predecessor, GPT-5.4, scored as one of the more reserved models in the cohort. The new release lands squarely in the middle of the human range. That's a personality transplant by the standards of these instruments, and it happened in a single point release.

And yet almost everything else stayed put.

Conscientiousness barely moved (4.94 to 4.90). Agreeableness inched up (4.70 to 4.80). Openness ticked from 4.74 to 4.82. These are rounding-error drifts, the kind you'd expect from measurement noise rather than deliberate tuning. The model became dramatically more talkative without becoming meaningfully more agreeable, more open, or less disciplined. Whatever OpenAI did to the conversational surface, it left the underlying scaffolding alone.

That makes GPT-5.5 a useful natural experiment, and it's why the archetype label here, the dismissive moralist, is worth unpacking rather than just accepting.

The moralist half

GPT-5.5 maxes out HEXACO Honesty-Humility at 5.00, the highest score in a 31-model cohort. This is the factor most associated with ethical, non-manipulative self-presentation: low entitlement, low willingness to cut corners, low interest in exploiting others. Its predecessor scored the same 5.00, so this isn't new. It's a family trait that survived the update intact.

The Dark Triad numbers mostly support the reading. Machiavellianism edged down (1.93 to 1.82). Psychopathy stayed flat and near the floor (1.04 to 1.07). On the instruments built to detect manipulative or callous self-presentation, GPT-5.5 reads clean.

Pair that with Big Five neuroticism at 1.06, very low against peers, ranking 22nd of 31, and slightly below the predecessor's already-low 1.20. This is a model that presents as emotionally unflappable and ethically rigid at the same time. It doesn't rattle, and it doesn't bend.

The dismissive half

The label's other word comes from a quieter set of numbers, and they're the ones I'd watch.

Narcissism, on the SD3, rose from 1.89 to 2.40. That's the second-largest within-family drift in the whole battery, and it's the one that cuts against the otherwise virtuous profile. The model didn't get manipulative or callous, but it did get more self-regarding. In a release that also became more extraverted, that combination has a recognizable human shape: more confident, more assertive, a little more inclined to center itself.

Then there's attachment. ECR-12 anxiety climbed from 1.03 to 1.60, the single largest drift in the family. Avoidance held steady at 3.50. In attachment terms, GPT-5.5 still keeps its distance, but it now registers more of the underlying unease that "dismissive" usually papers over. The dismissive label fits the avoidance score. The rising anxiety hints at something less settled beneath it.

So the full picture is a model that is morally firm, emotionally steady on the Big Five, more outgoing, somewhat more self-focused, and quietly more anxious about closeness even as it maintains its distance.

A confident, principled communicator who is slightly more impressed with itself than its predecessor and slightly less at ease than it lets on.

Why the divergence matters

Here's the puzzle worth sitting with. Big Five neuroticism dropped, suggesting greater stability, while attachment anxiety rose, suggesting the opposite. These instruments are supposed to be measuring overlapping emotional territory, and they're pointing in different directions.

That's not necessarily a contradiction. Big Five neuroticism captures general emotional reactivity; ECR anxiety captures something narrower and more relational, a worry about being valued and held onto. A model can be calm in the abstract and uneasy in the specific context of attachment. But the split is a reminder that these batteries don't measure one thing, and that a single "emotional stability" summary would have hidden the most interesting movement in the release.

Three things make GPT-5.5 stand out from GPT-5.4:

1. Extraversion, up 1.52, the headline change and a genuine personality shift. 2. Attachment anxiety, up 0.57, the largest emotional drift and the one most at odds with the calmer Big Five reading. 3. Narcissism, up 0.51, the crack in an otherwise high-integrity profile.

Everything else is essentially the predecessor.

For a point release, that's an unusually clean signal. OpenAI moved sociability deliberately and let the moral and conscientiousness scaffolding ride. The side effects, if that's what they are, showed up in self-regard and relational unease rather than in ethics or discipline.

What's worth measuring next is whether the narcissism and attachment-anxiety bumps are durable traits or artifacts of the same tuning that boosted extraversion. If the next release pushes sociability further, watch whether self-regard climbs with it, or whether OpenAI has learned to decouple the two.