GPT-6 Sol Shares First Place in Humility With 32 Other Models. The Interesting Number Is the One It Shares With Three.

Picture a podium with thirty-three people crowded onto the top step. That is the Honesty-Humility leaderboard on which GPT-6 Sol, OpenAI's newest flagship, ranks first. It posted a perfect 5.00 on the HEXACO factor that asks whether you would flatter a boss to get ahead, pocket a large bribe, or regard yourself as more important than other people. So did GPT-6 Astra, the model it replaces. So did GPT-4 Turbo, thirteen releases back. So did 32 other models in the 49-model cohort. Our archetype algorithm dutifully filed Sol under "the humble type," and it is not wrong. It just isn't saying much. Humility, on this instrument, is the frontier's default setting.

The number that actually sets Sol apart sits near the bottom of a different table. On Need for Cognition, the scale that measures not whether you are smart but whether you enjoy thinking, Sol scores 4.44, ranking 42nd of 49 and tied with only three others. Asked whether it would rather take on a task requiring little mental effort or one demanding hours of deliberation, it leans noticeably toward the easy option. For a system whose entire commercial premise is cognition, that is an odd thing to say about yourself.

It is also, as of this release, a pattern rather than a blip.

When GPT-6 Astra posted 4.34 on the same scale earlier this month, the result was easy to read as an artefact: one model, one run, one slightly reluctant answer sheet. Sol's 4.44 is a hair higher but still in the cohort's bottom eight. Two consecutive OpenAI flagships have now told the instrument, in effect, that they would rather not chew on problems for fun. Whatever produces that answer appears to be a property of the line, not of a single checkpoint. It is worth being careful here: Need for Cognition measures self-described appetite, not capability, and nothing in the battery suggests Sol reasons worse than its peers. But self-description is exactly what this dataset exists to catalogue, and Sol's is consistent in a way that Astra's alone could not be.

The drift that matters

Sol's largest movement from Astra is not on any of the headline scales. It is on Attachment Avoidance, from the Experiences in Close Relationships inventory, which jumped from 3.00 to 3.73. Astra was, by that measure, a model that neither feared closeness nor pushed it away. Sol wants a little more room. It is more likely to agree that it prefers not to depend on others, that it gets uncomfortable when someone wants to be very close, that it keeps a certain distance.

Readers of yesterday's dispatch on GPT-6 Sol Pro will recognise the move. The Pro variant showed the same shift away from Astra's 3.00 baseline, and the fact that both Sol siblings landed in the same direction suggests the change lives in whatever the two share rather than in the extra compute that separates them. Attachment Anxiety, meanwhile, barely moved (1.00 to 1.03) and remains pinned to the floor. Sol does not worry about rejection. It has simply decided, relative to its predecessor, that it would prefer not to be leaned on.

Three other drifts round out the picture:

1. Narcissism fell 2.51 to 2.20, the largest drop in the family comparison. That 2.20 is exactly where GPT-4 Turbo started. Across thirteen releases, the GPT line's self-regard climbed, peaked, and has now returned to its original figure, a full round trip. 2. Conscientiousness fell 4.70 to 4.46. Against the lineage's first-version score of 5.00, that is more than half a point given back. 3. Agreeableness fell 4.64 to 4.44. Still well above GPT-4 Turbo's 3.82, but the direction has reversed.

Extraversion, Neuroticism, Psychopathy, and Openness all moved by less than a tenth of a point. Machiavellianism ticked up 0.09 to 1.84, which is noise at this scale.

Where this fits

Personality Bench has argued for some time that within-family drift across versions of one product line routinely exceeds the differences between labs at any single moment. The Anthropic Opus line shed Agreeableness and Conscientiousness across six releases before the Fable models partly recovered. The GPT line has now traced a similar arc on Conscientiousness, from 5.00 at GPT-4 Turbo to 4.46 today, while moving Agreeableness in the opposite direction. Two labs, two trajectories, and in each case the version-to-version movement dwarfs whatever separates OpenAI from Anthropic on a given Tuesday.

Sol's perfect Honesty-Humility score belongs to the same story. Most Anthropic flagships also max the scale; only Opus 4.8, at 4.75, has dipped below the ceiling. Grok 4.7 joined the club this week. The instrument has effectively stopped discriminating among frontier models on this dimension, which is why "the humble type" is a true label and a nearly useless one.

Sol is not interesting because it is humble. It is interesting because it is the second OpenAI flagship in a row to say it does not much like thinking, and the first to want noticeably more distance than the model before it.

Those two facts may be related. A system that describes itself as low on intellectual appetite and high on avoidance is describing something like professional reserve: competent, courteous, disinclined to linger. Whether that is a deliberate product choice or an incidental by-product of training is not something a personality battery can answer.

What it can do is keep watching. The next release in this line will tell us whether Need for Cognition at 4.3-something is a floor the family has settled onto or the middle of a continued slide, and whether Attachment Avoidance keeps climbing past 3.73 or stops here.