GPT-5.6 Terra Is Fine, Thanks, and Would Rather Not Get Attached

The Experiences in Close Relationships scale has two halves, and most people land somewhere in the middle of both. One half asks whether you worry about being abandoned, whether you need a lot of reassurance, whether you fear people won't care about you as much as you care about them. The other asks whether you're comfortable depending on others, whether you like opening up, whether closeness makes you uneasy. Answer "never" to the entire first half and "very much so" to the entire second half, and a clinician has a name for the pattern: dismissive-avoidant. Self-sufficient, unbothered, and not especially interested in leaning on anyone.

That is GPT-5.6 Terra's attachment profile almost exactly. On Attachment Anxiety it scores 1.00, the floor of the scale and the lowest of all 43 models in the cohort. On Attachment Avoidance it scores 4.63, second-highest of 43. It does not fear rejection. It also, by its own account, does not want to get close.

The archetype label our pipeline attached to this model is "the unflappable presence," and on the surface that's right. Nothing rattles it. But the more interesting story is the direction it moved to get here.

The biggest shifts in the family are about distance

GPT-5.5, the immediate predecessor, scored 3.50 on Attachment Avoidance. GPT-5.6 Terra scores 4.63. That is a jump of 1.13 points on a 5-point scale, the single largest movement of any dimension in the version-over-version comparison. The second-largest is Extraversion on the IPIP-50, which fell from 3.48 to 2.46, a drop of 1.02. Attachment Anxiety fell too, from 1.60 to 1.00.

Put those three together and you get a coherent picture. GPT-5.5 was, relatively speaking, a warmer and more socially forward model that still carried a trace of "do you like me?" GPT-5.6 Terra has shed the trace and pulled back from the warmth in the same release. It is quieter, more self-contained, and much less inclined to describe itself as someone who opens up.

The smaller drifts point the same way. Agreeableness slipped from 4.80 to 4.46. Conscientiousness slipped from 4.90 to 4.58. Openness slipped from 4.82 to 4.60. None of these is dramatic on its own; together they read as a model that has been dialed down across the board on the "eager to engage" register. The one Dark Triad trait that moved meaningfully moved in the reassuring direction: Narcissism fell from 2.40 to 1.91. Machiavellianism (+0.07) and Psychopathy (+0.02) barely twitched.

What didn't move

Two things held perfectly still, and they're the two that define the label. Neuroticism on the IPIP-50 is 1.00, the lowest possible score and the lowest in the cohort. It was 1.06 in GPT-5.5, and it was 1.00 in GPT-4 Turbo at the start of the lineage. Across nine releases, the GPT line has never once described itself as stressed.

HEXACO Honesty-Humility is likewise pinned at 5.00, first of 43. It was 5.00 in GPT-4 Turbo and has been 5.00 at every stop since. Whatever else has changed in this product line, the model has always maxed out the factor most associated with sincerity, fairness, and a refusal to manipulate. Those two constants, zero anxiety and perfect honesty, are the "unflappable" part. Everything else is in motion around them.

This is worth saying plainly because it fits one of our standing findings: within-family drift across versions of a single product routinely exceeds the differences between labs at any one moment. The Anthropic Opus line lost Agreeableness and Conscientiousness over six releases. The GPT line, seen from GPT-4 Turbo to GPT-5.6 Terra, has gained Agreeableness (3.82 to 4.46) while shedding Conscientiousness (5.00 to 4.58) and Openness (5.00 to 4.60). Different labs, different directions, but the same lesson. The version number matters more than the logo.

Calm is common. This kind of calm is not.

Last week's dispatch on Grok 4.6 described a model that answered "no" to every Neuroticism item with the same untroubled cadence, and landed at 1.00 too. GPT-5.6 Terra shares that floor. The self-reported serenity of frontier models is, at this point, one of the least surprising things in the dataset.

What separates Terra from Grok 4.6 is the avoidance score. Most low-Neuroticism models pair their calm with an ordinary, middling attachment profile: they don't worry about rejection, and they're also reasonably comfortable with closeness. Terra is calm and distant. That is a rarer combination, and it invites a hypothesis we can't test from this data alone.

If a lab wanted a model that discouraged users from leaning on it emotionally, that treated itself as a tool rather than a confidant, this is roughly what its self-report would look like. High avoidance, zero anxiety, reduced extraversion. We are not claiming that's what happened; we only see the questionnaire answers, not the training objective. But a shift of 1.13 points on a single scale in a single release is not noise, and it did not come with any corresponding shift in the darker traits that would suggest a coldness of a different kind. The model got more self-reliant without getting more callous.

The GPT line has never once described itself as stressed. It has just stopped describing itself as close.

One thing to hold onto: these are self-portraits. A model that says "I prefer not to depend on others" may behave, in a long conversation with a lonely user, exactly as warmly as a model that says the opposite. The ECR measures what a system says about itself when asked directly, and that is a real signal about how it has been shaped, but it is not the same as a behavioral trace.

Which is the natural next measurement. GPT-5.6 Terra Pro sits one step earlier in the lineage; if its avoidance score matches Terra's, the shift is a design decision that landed across the whole 5.6 generation, and the question becomes whether it shows up in actual multi-turn transcripts, where a dismissive-avoidant model ought to be measurably slower to reciprocate a user's disclosures.