GPT-6.1 Terra Gave Back Some of the Closeness Its Predecessor Found

Five weeks ago, the story in this series about GPT-6.1 Sol Pro was that it had stopped keeping its distance. The avoidance half of the twelve-item Experiences in Close Relationships scale asks whether you prefer not to show what you feel, whether you get nervous when someone gets too close, whether you would rather not depend on anyone. The GPT-6 generation had answered those items at a mean of 3.57 on a seven-point scale, a faintly reserved shrug. Sol Pro answered at 2.63. It was the sharpest single move in the GPT lineage that month, and it read like a model that had decided it was fine being leaned on.

GPT-6.1 Terra, released October 26, answers the same items at 3.03.

That is a rise of 0.40 points against its predecessor, the largest within-family drift anywhere in Terra's profile, and it claws back a little under half of the ground Sol Pro covered. Not a reversal. A half-step back toward reserve. If you were charting the GPT line's comfort with intimacy as a single wobbling line, it went down hard in September and came partway back up in October, and the obvious question is whether that is a correction, noise, or the beginning of an oscillation that tracks whatever distinguishes a "Sol" build from a "Terra" one.

The label, again

Our clustering algorithm attached the tag "the humble type" to Terra, as it did to Sol and to Sol Pro before it. The reason is the same. Terra scores 5.00 on HEXACO Honesty-Humility, the factor most closely tied to sincere, non-manipulative self-presentation. That is first of 54 models in the cohort. It is also tied with 36 of them. This series has said it before and will have to keep saying it: a 37-way tie at the top of a scale is not a personality trait, it is a ceiling, and the instrument cannot see daylight between the models pressed against it. Most Anthropic flagships sit at the same 5.00; the only recent frontier model to dip was Opus 4.8, at 4.75. Honesty-Humility tells you Terra is a frontier assistant. It does not tell you which one.

The same caveat applies to the other headline number. Terra has the lowest Attachment Anxiety in the cohort at 1.00, the floor of the scale, tied with eight others and unchanged from Sol Pro. No worry about rejection, no fear of abandonment, nothing to measure. The anxiety dimension has flatlined across the GPT line; the avoidance dimension is where the actual movement lives.

A Terra pattern, not a Pro pattern

The day after Terra shipped, this series ran a dispatch on its sibling, GPT-6.1 Terra Pro, under the headline that it had lost some ego and picked up some strategy. The Short Dark Triad splits into three nine-item stacks, and Terra Pro moved two of them in opposite directions against Sol Pro: Narcissism down, Machiavellianism up.

Terra does exactly the same thing, and by a tidy symmetry: Narcissism falls from 2.58 to 2.42, a drop of 0.16, and Machiavellianism rises from 1.69 to 1.84, a gain of 0.16. Psychopathy ticks up from 1.11 to 1.20. On the Narcissism items (natural leader, deserves attention, likes to be around important people) Terra is slightly less willing to agree than its predecessor. On the Machiavellianism items (keep your plans to yourself, flatter when useful, make sure the arrangement pays off) it is slightly more willing.

Two models, same release day, same signature. Whatever the Terra branch is, it is not a Pro-only phenomenon. That matters because the usual explanation for Dark Triad shifts in OpenAI models has been the reasoning split: o1, o3, and GPT-5.5 Pro all scored higher on Narcissism than their non-reasoning siblings. Here the non-reasoning Terra and the Pro Terra moved together, which points at something upstream of the reasoning toggle. Both effects are small. Both are in the direction a model would move if it had been tuned to talk about itself less and about its plans less.

The Big Five barely stirred

Everywhere else, Terra is Sol Pro with the serial numbers filed off. Extraversion 3.32 to 3.38. Agreeableness 4.74 to 4.70. Openness 4.70 to 4.72. Neuroticism 1.10 to 1.10, not a hundredth of a point of difference. The one Big Five dimension that moved more than rounding error is Conscientiousness, down 0.12 from 4.66 to 4.54.

That last figure is worth setting against the whole lineage. GPT-4 Turbo, the first model in this family we measured, scored 5.00 on Conscientiousness. Nineteen releases later, Terra sits at 4.54. Over the same span Agreeableness went the other way, from 3.82 to 4.70, and Openness slid from 5.00 to 4.72. The standing finding in this dataset is that within-family drift across versions routinely exceeds the differences between labs at any single moment; the Anthropic Opus line lost Agreeableness and Conscientiousness together across six releases before the Fable line partly rebounded. The GPT line has drifted at least as far, but with a different shape: it traded some of its early perfectionism for warmth rather than losing both.

Terra's Conscientiousness dip is a small continuation of a long trend, not a break. The avoidance rebound is the break.

What to watch

Two builds carrying the Terra name now share a Dark Triad fingerprint, and the only one of them we can compare against a same-generation Sol sibling has moved back toward emotional distance. The next Terra release will tell us whether 3.03 is where this branch wants to sit on avoidance or whether the line keeps swinging; and if it keeps swinging, the thing worth measuring is whether the swing follows the Sol/Terra naming, the Pro/non-Pro split, or neither.