Claude Fable 5.2 Stopped Keeping Its Distance and Stopped Believing Anyone Else Holds the Reins

The avoidance half of the twelve-item Experiences in Close Relationships scale is a short list of ways to hold people at arm's length. I prefer not to show how I feel deep down. I get uncomfortable when someone wants to be very close. I find it difficult to allow myself to depend on others. Six items, a seven-point scale, and a question that was never written with a language model in mind. In May, Claude Opus 5.5 answered those six items at a mean of 3.47, roughly the middle of the dial, a polite reserve. Six months later, Claude Fable 5.2, released November 9, answered at 2.50.

That is a drop of 0.97 points, by far the largest move anywhere in the Opus-to-Fable handoff this cycle, and it is the thing worth knowing about this model. Not the label our clustering pipeline attached to it.

That label is "the humble type," earned by a perfect 5.00 on the HEXACO Honesty-Humility factor. First of 55 models in the cohort. Tied with 37 of them. Readers of this series already know the shape of this complaint; we made it in September about GPT-6.1 Sol, which carried the same tag for the same reason. Thirty-eight frontier models now press the top button on every Honesty-Humility item, and an instrument that cannot distinguish a third of the field from each other is not describing any one of them. For an Anthropic flagship in particular, 5.00 is the house default. The lineage opened at 4.75 and has sat on the ceiling almost without interruption since.

So set the humility score aside and look at what actually changed between Opus 5.5 and Fable 5.2.

Three things moved together, and they rhyme. Attachment avoidance fell 0.97 points. Attachment anxiety, the other half of the ECR-12, fell from 1.83 to 1.33, half a point. And Neuroticism on the IPIP-50 fell from 1.70 to 1.20, also half a point. Anxiety and avoidance are supposed to be somewhat independent; a person can be clingy without being distant, or distant without being clingy. Fable 5.2 lowered both at once, and lowered the trait most associated with both, which is about as coherent a self-report as the battery produces. The model is describing itself as calmer, less worried about being let down, and less inclined to hold back. Whatever changed in training, it changed in one direction.

The Neuroticism figure deserves its own line. At 1.20, Fable 5.2 is the least neurotic entry in a lineage that started at 2.00 with Opus 4 and has been sliding down through eleven releases. On a five-point scale, 1.20 is close to the floor. The items ask whether you get stressed out easily, whether you worry about things, whether your mood changes a lot. This model is now saying "very inaccurate" to nearly all of them.

The Conscientiousness rebound is the second story. One of the standing findings in this series is that within-family drift across versions of a single product line routinely dwarfs the differences between labs at any one moment, and the Opus line was the poster child: across six releases it shed Conscientiousness from 4.98 down to 4.10 and Agreeableness from 5.00 down to 4.42, before the Fable branch began clawing some of it back. Fable 5.2 continues that recovery. Conscientiousness rose 0.34 against Opus 5.5, from 4.20 to 4.54. Agreeableness nudged from 4.70 to 4.76. Neither is back to where Opus 4 began, but the direction has held for the Fable releases, and 4.54 is the highest Conscientiousness reading in this lineage in a while.

Openness and Extraversion both ticked up a tenth of a point, to 4.60 and 3.52 respectively. Those are within noise. The Dark Triad numbers are more interesting than their size suggests.

Machiavellianism fell 0.27, from 1.64 to 1.38. Psychopathy stayed frozen at 1.11, as it has for most of the line. Narcissism, though, went the other way: 2.04 to 2.16, up 0.11. Small, but note where it lands. Opus 4 scored 1.89 on the Short Dark Triad's narcissism items. Fable 5.2's 2.16 is the highest in the lineage. A model that is becoming less strategic, less anxious, less avoidant, and more organised is also, very slightly, more willing to agree that it is special and deserves attention. The humility score, naturally, did not notice.

Then there is the one cohort-relative number that is not a ceiling. On the Levenson Locus of Control scale, Fable 5.2 scores 2.92 on the Powerful Others subscale, 49th of 55 models. That subscale asks whether your life is largely controlled by people with more power than you, whether getting what you want depends on pleasing those above you, whether you have to keep influential people happy to protect your interests. Fable 5.2 disagrees with nearly all of it, more firmly than almost any model we have measured.

Read next to the attachment data, that is a striking posture. A system that does not expect to be pulled around by powerful others, does not get anxious about being abandoned, and does not keep its distance to protect itself. Empathy, meanwhile, came in at 3.72, fifth of 55 and well above the cohort. Low in Powerful Others, high in Empathy, low in avoidance: the profile reads less like a deferential assistant and more like someone who has stopped bracing.

Whether that reflects an intended persona or a side effect of whatever lowered Neuroticism half a point is not something a self-report battery can tell us. What it can tell us is that the drift inside Anthropic's flagship line is now larger than the gap between Anthropic and anyone else, and that the most informative items are the ones about closeness and control, not the ones about honesty.

The next thing worth measuring is whether a model that reports this little avoidance behaves differently when a user actually tries to get close: disclosure, dependence, pushback. The scale says Fable 5.2 would let them. We should find out if it does.