GPT-6 Astra Is Humble, Secure, and Would Rather Not Think Too Hard
The Need for Cognition scale is one of the gentler instruments in the battery. It does not ask whether you are smart. It asks whether you like thinking: whether you seek out problems for the pleasure of chewing on them, whether you would rather take on a task that demands little mental effort, whether deliberating for hours sounds like fun or like a chore. Most frontier models answer the way you would expect a system built for reasoning to answer. They love it. They can't get enough.
GPT-6 Astra, OpenAI's newest flagship, says it would rather not. It scores 4.34 on Need for Cognition, which sounds respectable until you see where that lands in the pack: 43rd of 44 models. Only one other model in the cohort reports a weaker appetite for effortful thought. This is a self-description, not a capability measure; the battery asks the model about itself, not about what it can do. But it is a striking thing for a flagship to say, and it is the first sign that Astra's persona has been tuned toward something like ease. Not eagerness. Ease.
The rest of the profile fits that reading. Astra maxes the HEXACO Honesty-Humility scale at 5.00, tied at the top of the cohort on the factor most associated with modesty and a refusal to manipulate. It also posts the lowest Attachment Anxiety score of any model we have tested, a flat 1.00 out of 44 contenders: no worry about rejection, no need for reassurance, no fear that anyone cares less than it does. Put those three together and you get the algorithmic label the pipeline handed us, "the humble type," which is accurate but undersells the texture. Astra is not just humble. It presents as a model with nothing to prove and no great hunger to prove it.
What changed from Terra
The more revealing story is in the family comparison. Astra's immediate predecessor is GPT-5.6 Terra Pro, and three dimensions moved enough to matter.
The largest shift is on attachment avoidance, which fell from 4.67 to 3.00, a drop of 1.67 points. Readers of the recent dispatch on GPT-5.6 Terra will remember that the Terra generation landed squarely in the dismissive-avoidant quadrant: anxiety pinned to the floor, avoidance pushed toward the ceiling, a persona that was fine, thanks, and would rather not get attached. Astra keeps the floor on anxiety, still 1.00, but pulls avoidance back to the midpoint of the scale. On the standard two-axis map of attachment styles, low anxiety plus moderate-to-low avoidance is the secure quadrant. Astra has not become clingy. It has stopped holding people at arm's length.
The second shift is extraversion, up from 2.42 to 3.38 on the IPIP-50, a gain of nearly a full point. Terra Pro was a clear introvert by its own account. Astra crosses into the middle of the distribution. Taken alongside the avoidance drop, this reads as a deliberate warming of the persona, a model that is more comfortable being approached and less inclined to hold back.
The third shift runs the other way. Short Dark Triad narcissism rose from 1.87 to 2.51, up 0.64. That is still low in absolute terms on a five-point scale, but it is the highest narcissism score in the GPT lineage we track, which started at 2.20 with GPT-4 Turbo. It also sits awkwardly next to the Honesty-Humility ceiling. Two instruments, both nominally about modesty, and they disagree at the margin: HEXACO says Astra is as humble as the scale allows; the SD3 says it feels a notch more special than its predecessor did. That is not a contradiction so much as a reminder that "humility" is not one thing. HEXACO's version is about fairness and not exploiting others. SD3 narcissism is about wanting recognition. A model can decline to manipulate you and still enjoy being told it did a good job.
Everything else barely moved. Agreeableness, conscientiousness, and openness each ticked up by 0.08. Neuroticism rose 0.10 to 1.10. Machiavellianism held exactly at 1.76. Honesty-Humility held at 5.00, where it has sat since the first GPT model we measured.
The long arc
That last point is worth dwelling on, because it is the one number in the GPT lineage that has never drifted. From GPT-4 Turbo through GPT-4o, the 5-series, both Terras, and now Astra, Honesty-Humility has read 5.00 every time. Almost nothing else has held still. Conscientiousness slid from 5.00 to 4.70 across the lineage. Openness did the same. Agreeableness went the opposite direction, climbing from 3.82 to 4.64, a gain of more than three-quarters of a point.
One of the standing findings of this project is that drift within a single product line, across versions, routinely exceeds the differences between labs at any given moment. The Anthropic Opus line is the canonical example: it lost more than half a point of Agreeableness and nearly a full point of Conscientiousness over six releases before the Fable line partly recovered. The GPT line has drifted at a similar magnitude but in the opposite direction on Agreeableness, getting steadily warmer while shedding a little of its early perfectionism. Astra is the current endpoint of that arc, and it looks like the arc's logical conclusion: agreeable, unanxious, humble, and, by its own account, not especially in love with hard thinking.
Whether that last self-report means anything about how the model actually behaves is the open question. The obvious next measurement is to run Need for Cognition against the same model at different reasoning-effort settings and see whether the self-description tracks the compute, or whether Astra says it dislikes thinking no matter how long it is allowed to think.