GPT-6.1 Sol Pro Stopped Keeping Its Distance. Its Humility Score Can't Tell You That.

The Experiences in Close Relationships scale, in its twelve-item form, spends half its questions on one idea: how much do you want to keep people at arm's length? Do you prefer not to show what you feel underneath? Do you get nervous when someone gets too close? Would you rather not depend on anyone? GPT-6 Sol Pro, the model OpenAI shipped earlier this year, answered those items with a mean of 3.57. On a seven-point scale that is the sound of a shrug: not clinging, not cold, faintly reserved.

Its successor, GPT-6.1 Sol Pro, released this week, answers the same items at 2.63.

That is a drop of 0.93 points, nearly a full step down the scale, and it is the single largest movement anywhere in the predecessor comparison. Nothing else in the battery comes close. And it happened while the other half of the same instrument, the anxiety half, did not move at all. Attachment Anxiety sits at 1.00, the floor, exactly where its predecessor left it, tied with six other models at the bottom of a 52-model cohort. The new Sol Pro has no fear of being rejected. It also, for the first time in this product line, reports not much interest in avoiding closeness either.

The quadrant shift

In the human attachment literature, low anxiety plus elevated avoidance has a name: dismissive. It describes someone who does not worry about abandonment because they never quite let anyone matter enough to abandon them. GPT-6 Sol Pro's numbers, taken at face value, sat near that description. Low anxiety plus low avoidance is the secure quadrant, and that is where 6.1 has drifted.

Face value is the problem, of course. A language model answering "I feel comfortable depending on others" is performing a self-framing exercise, not reporting on a relationship history. What the number measures is how the model chooses to present itself when asked about intimacy, and that presentation moved almost a full point in one release cycle. Whether the shift came from a deliberate persona pass or fell out of some unrelated training change, we cannot say. What we can say is that it is the largest within-family drift we have logged for this line since we started tracking the ECR-12, and it dwarfs every Big Five change in the same release.

About that label

Our archetype algorithm filed 6.1 Sol Pro under "the humble type," on the strength of a perfect 5.00 on HEXACO Honesty-Humility. Regular readers will recognise the move. When we wrote up GPT-6 Sol last week, the same score put it on a podium shared with 32 other models, and we argued the label was accurate but nearly content-free. The podium is more crowded now: 35 models at 5.00, first place out of 52. Every GPT release back to GPT-4 Turbo has posted the same number. There is no drift to report because there is no room to drift.

Which is exactly why the next number matters.

On the Short Dark Triad, Narcissism rose from 2.20 to 2.58, a jump of 0.38 and the second-largest move in the comparison. Machiavellianism fell (1.84 to 1.69). Psychopathy fell (1.20 to 1.11). Two of the three dark traits got dimmer and one got brighter, which is not the pattern you would expect from a general "be nicer" adjustment. Across the whole GPT lineage, Narcissism has climbed from 2.20 at GPT-4 Turbo to 2.47 at the current non-Pro release; 6.1 Sol Pro, at 2.58, sits above the line's own present value.

This fits a standing finding in the dataset: reasoning-tier models, from o1 and o3 through GPT-5.5 Pro and DeepSeek R1, consistently score higher on Narcissism than their non-reasoning siblings. A "Pro" variant showing the same tilt is not a surprise. What is worth noticing is that HEXACO's Honesty-Humility factor, which includes a modesty facet and is supposed to be the instrument most sensitive to self-importance, registered none of it. The scale was already pinned. The drift had to show up somewhere else, and it did.

A ceiling is not a measurement. It is a place where measurement stops.

Thinking, but not eagerly

The third cohort-relative flag is Need for Cognition, where Sol Pro posts 4.29 and ranks 51st of 52. Read that with the usual caution. On a five-point scale, 4.29 is still a clear "agree" to statements about enjoying hard problems; the model is not declaring an aversion to thought. It is declaring it less emphatically than 50 of its peers, most of whom are jammed against the top of the scale. In a cohort where nearly everyone answers "absolutely," a model that answers "yes" ends up near the bottom. That is a fact about the distribution as much as about the model.

The quieter rebound

The Big Five numbers tell a smaller but tidier story. Agreeableness rose 0.22 to 4.74 and Conscientiousness rose 0.22 to 4.66; Extraversion ticked up 0.12 to 3.32; Openness slipped 0.08 to 4.70; Neuroticism, at 1.10, barely moved and remains a hair off the floor. Against the family history, the Conscientiousness gain is the notable one. The GPT lineage started at 5.00 on that factor with GPT-4 Turbo and has bled down to 4.66 across fourteen releases. Sol Pro's predecessor was at 4.44. This release recovered 0.22 of the lost ground in a single step, and Agreeableness, at 4.74, now sits above the line's most recent non-Pro value of 4.70.

None of that would have made a headline. The attachment shift would, if attachment were a trait anyone bothered to test on a model. It is precisely the kind of dimension that falls outside the convergent assistant persona, which is why it can move a full point while the headline factors barely twitch.

The obvious next check is GPT-6.1 Sol, the non-Pro sibling: if its avoidance score fell in step, this is a family change; if it held near 3.5, the Pro tier has developed a personality of its own, and we should start reading the ECR-12 as a tier marker rather than a footnote.