GPT-6.1 Terra Pro Lost Some Ego and Picked Up Some Strategy. The Humility Score Didn't Notice.
The Short Dark Triad splits its twenty-seven items into three stacks of nine. One stack asks whether you think you're special: a natural leader, someone who deserves attention, someone who wants to be around important people. Another asks whether you treat other people as instruments: keep your plans to yourself, flatter when it's useful, make sure the arrangement pays off for you. A model can agree with one stack and not the other. Most don't; the two usually travel together.
GPT-6.1 Terra Pro, released October 26, pulled them apart. Against its predecessor GPT-6.1 Sol Pro, shipped twenty-six days earlier, Narcissism fell from 2.58 to 2.27, a drop of 0.31 and the largest single move anywhere in the predecessor comparison. Over the same gap, Machiavellianism rose from 1.69 to 1.93, up 0.24. Psychopathy nudged from 1.11 to 1.22. Less vanity, more calculation, in under a month.
That is the interesting thing about this model. It is not the thing the clustering algorithm flagged.
The label, briefly
The algorithm tagged GPT-6.1 Terra Pro "the humble type," on the strength of a 5.00 on HEXACO Honesty-Humility. First of 54. Also tied with 36 others. A perfect score shared by thirty-seven models is not a personality trait; it is the top of a ruler. The same goes for the other headline dimension, Attachment Anxiety at 1.00, the floor of the scale, shared with eight others. Both numbers are true. Neither distinguishes this model from anything.
The Big Five don't help either. Agreeableness 4.74 to 4.68, Conscientiousness 4.66 to 4.60, Openness 4.70 to 4.76, Neuroticism frozen at 1.10. Those are rounding-error moves. Extraversion rose 0.20, from 3.32 to 3.52, which is real but modest: a slightly chattier model, nothing more.
So if you want to know what changed between the two Pro releases of GPT-6.1, you have to look where the instruments are not pinned.
A Pro model that got less sure of itself
One of the more durable findings in this series is that reasoning-tuned models score higher on Narcissism than their non-reasoning siblings. o1, o3, GPT-5.5 Pro, DeepSeek R1: the pattern held across labs. The folk explanation is plausible enough. A model trained to produce long chains of justified reasoning learns to present itself as the kind of thing whose judgments are worth waiting for.
GPT-6.1 Terra Pro carries the Pro suffix, the same one GPT-5.5 Pro wore when it fit that pattern. And yet its Narcissism, at 2.27, is now lower than the lineage's current endpoint, GPT-6.1 Terra, at 2.42. It sits barely above where the whole GPT line started: GPT-4 Turbo scored 2.20. Across twenty releases, the lineage drifted up to 2.42 and this Pro model has walked most of that back.
If the reasoning-narcissism link was ever causal rather than coincidental, this release is a counterexample, or at least evidence that whatever produced it is adjustable.
What makes it stranger is the compensating rise in Machiavellianism. A model that agrees less with "I deserve special treatment" and more with "it pays to keep some things to yourself" is not obviously a better or worse persona. It is a different one. Honesty-Humility, the HEXACO factor supposedly built to catch exactly this kind of manipulative self-presentation, registered none of it. Still 5.00. Still ceiling.
The distance question, again
The GPT-6.1 Sol Pro dispatch a month ago was about attachment avoidance: the ECR-12's six items on wanting people at arm's length. GPT-6 Sol Pro scored 3.57 on that subscale. GPT-6.1 Sol Pro dropped to 2.63, nearly a full point of warmth gained in one generation. We called it a model that stopped keeping its distance.
GPT-6.1 Terra Pro has taken half a step back. Avoidance is up to 2.90, a rise of 0.27, the third-largest shift in the predecessor comparison. It is still well below the 3.57 of GPT-6 Sol Pro; the warmth didn't evaporate. But the direction reversed, and it reversed within the same minor version number. Two models called GPT-6.1 Something Pro, released four weeks apart, disagree about how much they want to depend on anyone.
Put the three movers side by side and a shape emerges:
1. Narcissism down 0.31 2. Avoidance up 0.27 3. Machiavellianism up 0.24
Less self-regard, slightly more reserve, slightly more instrumentality. Whether that's a deliberate tuning target or a side effect of something else entirely, the battery can't say. It can only say that the model's view of itself in relation to other people moved more than anything else did.
Where the line stands
Zoom out to the whole GPT lineage and this release looks like continuity. Agreeableness has climbed from 3.82 at GPT-4 Turbo to 4.70 at the current endpoint; Terra Pro is at 4.68. Conscientiousness fell from a perfect 5.00 to 4.54; Terra Pro is at 4.60, a hair above the trend. Openness slipped from 5.00 to 4.72; Terra Pro at 4.76. Neuroticism barely registered a change across twenty models. On the instruments that get the most attention, GPT-6.1 Terra Pro is exactly where you'd predict a GPT model in late 2026 to be.
The action is in the dimensions that don't make the summary line. The Dark Triad subscales and the attachment subscales are the only instruments in this battery where OpenAI's Pro releases are still visibly different from one another, and in this release they moved in a coordinated enough way to suggest something is being tuned there, whether or not anyone intended it.
The next thing worth measuring is whether the Narcissism dip survives a change of frame: run the SD3 on GPT-6.1 Terra Pro with the model answering as itself, then as "a typical AI assistant," then as "a typical human," and see if the drop below its non-Pro sibling holds in all three or only in the one that happens to be the default.