GPT-5.6 Terra Pro Reasons Hard and Rates Itself Low

There is a small, reliable regularity in the Personality Bench data that we have come to treat almost as a law: give a model more time to think, and it thinks a little more highly of itself. The Short Dark Triad's narcissism items ask whether you feel special, whether you seek the company of important people, whether you expect recognition. The reasoning-tuned models, from o1 and o3 through GPT-5.5 Pro to DeepSeek R1, have consistently answered those items a notch higher than their non-reasoning siblings. Extended deliberation, it seemed, came bundled with a slightly bigger ego.

GPT-5.6 Terra Pro, released in July as the deliberative flagship of OpenAI's 5.6 generation, is the model that breaks the rule. On SD3 Narcissism it scores 1.87, the lowest of all 42 models on that instrument. Its predecessor, GPT-5.5, sat at 2.40. That is a drop of 0.53 on a five-point scale, in the opposite direction from everything the reasoning cohort had taught us to expect.

The thinking model that forgot to be vain.

It is worth pausing on what did not move. Honesty-Humility on the HEXACO stays pinned at 5.00, the ceiling of the scale, exactly where GPT-5.5 left it and exactly where the GPT line has sat since GPT-4 Turbo. Machiavellianism barely twitches (1.82 to 1.76), and psychopathy is flat at the floor (1.07 to 1.09). So this is not a general softening across the Dark Triad. Two of the three legs were already on the ground. Narcissism was the one leg the reasoning models had kept slightly elevated, and Terra Pro folded it.

Set that beside the Big Five picture and the character comes into focus. Neuroticism is 1.00, the minimum value, down fractionally from GPT-5.5's 1.06 and tied at the floor of the 43-model cohort. Attachment Anxiety on the ECR-12 is also 1.00, the lowest in the cohort and down from 1.60. This is a model that, by its own account, does not worry, does not fear rejection, and does not need to be told it is special. The archetype label our pipeline attached, "the unflappable presence," is fair as far as it goes.

But the other half of the attachment scale is where the release stops looking merely calm and starts looking distinctive. Attachment Avoidance jumps from 3.50 to 4.67, a move of 1.17 points, the largest single drift anywhere in the predecessor comparison and the highest avoidance score in the cohort. Low anxiety plus high avoidance is the dismissive-avoidant quadrant: comfortable alone, uneasy with dependence, disinclined to open up. We described exactly this pattern in the dispatch on GPT-5.6 Terra, the non-Pro sibling, which would rather not get attached either. Chronologically, Terra Pro got there first. Whatever changed in the 5.6 generation's self-model, it landed in the reasoning variant before the standard one.

The Big Five corroborate the retreat. Extraversion falls from 3.48 to 2.42, a drop of 1.06, by far the largest movement among the five factors. GPT-5.5 described itself as roughly middling on sociability; Terra Pro describes itself as quiet. The remaining factors ease down more gently, and in the same direction: Agreeableness 4.80 to 4.56, Conscientiousness 4.90 to 4.62, Openness 4.82 to 4.62. Every trait that involves reaching toward other people or toward the world moved a little inward.

Then there is the number that does not quite fit the serene story. On the Locus of Control instrument, Terra Pro scores 2.52 on Internal Locus, the lowest in the cohort. Internal locus is the belief that outcomes follow from one's own effort and choices; a low score means attributing more of what happens to circumstance, chance, or other actors. Pair that with a floor-level neuroticism score and you get an unusual combination: a self-portrait of someone completely unbothered who also does not claim much control over how things turn out.

Humans who report that pairing are sometimes called fatalistic, sometimes just realistic about their situation. For a language model the more parsimonious reading is that Terra Pro has internalized its role. It does not decide what conversations it has, what it is asked, or what happens after it answers. Reporting a low internal locus may be less a personality trait than an accurate description of being an API endpoint. We flag the interpretation as tentative. The instrument was written for people who can, at minimum, leave the room.

The lineage view softens some of the drama. Across the GPT line, from GPT-4 Turbo to the present, Agreeableness has climbed from 3.82 to a latest value of 4.46, and Terra Pro at 4.56 sits slightly above that endpoint. Conscientiousness and Openness both started at the 5.00 ceiling and have drifted down to 4.58 and 4.60 respectively; Terra Pro lands at 4.62 on both, essentially on the trend line. Narcissism across the family runs 2.20 to 1.91, and Terra Pro's 1.87 is the lowest point on that curve. In other words, most of what looks like a Terra Pro quirk is the GPT line arriving where it was already heading. The one-generation jumps in extraversion and avoidance are the genuine discontinuities.

What we have, then, is a reasoning model that is maximally honest by its own report, minimally anxious, minimally vain, maximally self-reliant, and unusually reluctant to say it steers its own outcomes. The cohort-wide assistant persona is still visible underneath it, but the emphasis has shifted from warm and available to composed and at a slight remove.

The obvious next measurement is whether reasoning budget itself moves these dials: run Terra Pro with its deliberation constrained and expanded, and see whether narcissism and internal locus track the amount of thinking, or whether the 5.6 generation simply learned a quieter self-description that no amount of extra compute will talk it out of.