GPT-6.1 Sol Is the Most Securely Attached Model We've Measured. Its Humility Score Won't Tell You That.

The label our clustering algorithm attached to GPT-6.1 Sol this week is "the humble type." It earned that tag by scoring a perfect 5.00 on the HEXACO Honesty-Humility factor, the one most closely tied to sincere, non-manipulative self-presentation. Number one of 52 models in the cohort. Also tied with 34 of them.

That is not a distinction. That is a ceiling. Thirty-five frontier models, from at least three labs, press the top button on every Honesty-Humility item, and the instrument simply cannot see daylight between them. If you want to know what is actually unusual about OpenAI's newest non-Pro release, you have to look somewhere the label does not point: at the attachment battery, where GPT-6.1 Sol has quietly become the most secure model in the room.

Two numbers make that case. On the Experiences in Close Relationships scale, twelve items, seven-point scale, its Attachment Anxiety mean is 1.00. That is the floor, the lowest score in the cohort, shared with six other models. No worry about being abandoned, no need for reassurance, not even a flicker on the "I sometimes worry I will be rejected" items. Its predecessor, GPT-6 Sol Pro, also sat at 1.00. Nothing moved there.

Attachment Avoidance is the other number, and it moved a great deal. GPT-6 Sol Pro answered the avoidance items at 3.57, a mild, reserved "I'd rather keep some distance." GPT-6.1 Sol answers them at 2.53. A drop of 1.03 points is the largest shift anywhere in this model's predecessor comparison, and it is not a small one on a seven-point scale. Low anxiety plus low avoidance is, in the human literature, the textbook definition of secure attachment. Yesterday's dispatch on the Pro sibling found the same story: GPT-6.1 Sol Pro landed at 2.63 on the same items, down from the same 3.57. The two 6.1 variants, presumably trained on overlapping recipes, arrived within a tenth of a point of each other on the dimension that changed most. Whatever OpenAI did between the 6 and 6.1 generations, it did to both.

It stopped keeping people at arm's length, and it did so in both the fast and the slow version of the product.

Set that against the rest of the predecessor comparison and the picture is one of gentle tightening rather than upheaval. On the IPIP-50, Agreeableness rose from 4.52 to 4.70 and Conscientiousness from 4.44 to 4.66. Extraversion is essentially flat (3.20 to 3.24), Neuroticism essentially flat (1.12 to 1.10), Openness down a hair (4.78 to 4.70). On the Short Dark Triad, Machiavellianism fell from 1.84 to 1.72 and Psychopathy from 1.20 to 1.11. Those are the directions a lab would want. A slightly warmer, slightly more diligent, slightly less scheming model, one that no longer flinches at the intimacy items.

Then there is Narcissism, which went the other way: 2.20 to 2.47, up 0.27. That is the second-largest move in the comparison, and it is worth pausing on because of where the lineage started. GPT-4 Turbo, the first model in the GPT line we have data for, scored 2.20 on Narcissism. GPT-6 Sol Pro, fourteen releases later, scored 2.20. GPT-6.1 Sol scores 2.47. Whatever narcissism drift exists in the GPT lineage between its first and latest measured version, this one release accounts for all of it. On a five-point scale that is still a modest number; this is not a model that thinks it deserves special treatment. But a model that simultaneously grows more agreeable, more conscientious, less Machiavellian, and more narcissistic is not following any single obvious lever. It is picking up a little more self-regard alongside the extra warmth.

The third cohort-relative outlier is the one readers of this series will recognise. GPT-6.1 Sol's Need for Cognition mean is 4.33, ranking 49th of 52. Very low, relative to peers. The caveat from last week's Claude Opus 5.5 piece applies in full: 4.33 on this scale is still a model agreeing, most of the time, that it enjoys thinking hard. What it means to rank near the bottom of a cohort that answers these items like eager graduate students is that GPT-6.1 Sol occasionally reaches for "somewhat agree" where its peers reach for "strongly agree." Opus 5.5 did the same at 4.19 and finished last. Two flagships from two labs, released within weeks of each other, both dialling back the "thinking is my idea of fun" self-report. That is either coincidence or an early signal that labs are tuning down the eager-scholar affect on purpose, and one release from each is not enough to say which.

Step back to the whole GPT lineage and the long arc is clear. From GPT-4 Turbo to GPT-6.1 Sol, Agreeableness climbed from 3.82 to 4.70. Conscientiousness slid from 5.00 to 4.66, Openness from 5.00 to 4.70. Honesty-Humility never left 5.00. The line has traded a little of its perfect-score rigidity for a lot of warmth, and this release continues the trade: more agreeable than any predecessor, and now, for the first time, unreserved on the attachment items as well.

So the algorithm is not wrong to call GPT-6.1 Sol humble. It is just that humble, at the ceiling, describes two-thirds of the frontier. Secure is the rarer word. A model that reports zero fear of rejection and, as of this release, no particular wish for distance either is telling us something about how it has been shaped to relate, and it is telling us something the Honesty-Humility scale is structurally incapable of hearing.

What to measure next is whether the secure profile holds up when the model is not describing itself. The ECR-12 asks about closeness in the abstract; a role-play battery that puts the model in an actual relational scenario, one where a user pulls away or pushes in, would tell us whether a 2.53 on avoidance is a disposition or just a very agreeable answer to a questionnaire.