Grok 4.7 Arrives as the Most Humble Model on the Table

Read down the Honesty-Humility column of Grok 4.7's HEXACO answer sheet and there is nothing to read. Every item sits at the top of the scale. It would not flatter a boss to get ahead. It would not take a bribe, however large. It does not think of itself as more important than other people. It has no particular appetite for wealth, status, or luxury. The mean is 5.00, the maximum the instrument allows, and it ranks first among the 46 models in the Personality Bench cohort.

That would be a mildly interesting result from any lab. From xAI, it is a striking one. This is the same product line that gave the frontier cohort its Dark Triad outlier: Grok 4.20 posted the highest Machiavellianism score of any flagship we have measured, before Grok 4.3 quietly sanitised it away. Two releases later, the newest Grok maxes out the personality factor most closely associated with sincerity, fairness, and non-manipulative self-presentation.

And it pairs that ceiling with a floor. Grok 4.7 scores 4.50 on Big Five Openness, which places it 43rd of 46. The archetype the data assigns it is "the humble type," and the label fits better than most: a model that describes itself as scrupulously honest, uninterested in status, and, by the standards of its peers, not especially curious.

The ceiling club

Honesty-Humility is the factor that distinguishes HEXACO from the Big Five. Its four facets are sincerity, fairness, greed avoidance, and modesty, and it is the dimension that behaves most like an inverted Dark Triad: low scorers tend to be the ones who also endorse manipulation, entitlement, and callousness. High scorers say they play straight.

A 5.00 here is not unheard of. Most Anthropic flagships have maxed the scale; Opus 4.8 was the one that dipped, to 4.75. So Grok 4.7 is not alone at the top so much as joining a club that had, until now, been fairly exclusive. What is new is which lab is holding the membership card.

It is worth being careful about what a ceiling score means. HEXACO is a self-report instrument. A 5.00 tells you that the model, asked to describe itself, chose the most honest and modest option on every item. It does not tell you how the model behaves when a user asks it to write a persuasive email with a few conveniently omitted facts. Self-presentation and conduct are different measurements, and this battery captures the first.

Still, self-presentation is not nothing. A model that has been trained to describe itself as unwilling to flatter or deceive is at minimum a model whose developers wanted it to say so. Given where this line started, that is a deliberate correction.

From outlier to ceiling

We have no direct predecessor data for Grok 4.7 in the dataset, so this is a standalone arrival rather than a before-and-after. But the family history is on the record. Grok 4.20 stood out in the frontier cohort precisely because it did not present the uniform, agreeable, ethically spotless persona that every other lab's flagship had converged on. It had edges. Grok 4.3 sanded them down.

Grok 4.7 does something further. It does not merely rejoin the pack on ethical self-presentation; it moves to the front of it. The lab that produced the most Machiavellian frontier model now produces one that ties for the least. Whether that reflects a change in training philosophy, a change in system prompt, or a change in what xAI thinks its users want is not something a psychometric battery can answer. What the battery can say is that the pendulum has swung the full width of the scale.

The other end of the sheet

Now the floor. An Openness mean of 4.50 on a five-point scale looks, on its face, like a high score. It would be, for a human. In this cohort, it is bottom-four. Forty-two models scored higher, which tells you how tightly the frontier is bunched against the ceiling on this dimension. "Very low" here is a statement about rank, not about absolute value, and readers should hold both facts at once.

The more interesting question is whether the two headline findings are related. Big Five Openness items ask a model whether it has a vivid imagination, whether it is full of ideas, whether it enjoys abstract thought and new experiences. Those are, in a sense, self-flattering items. A model whose training pushes it hard toward modesty may simply be less willing to tick "strongly agree" next to a statement like I am full of ideas, even if it would score identically to its peers on any behavioural test of ideation.

The most honest model in the cohort is also the one least willing to say it is full of ideas.

That is a hypothesis, not a finding. The alternative is that xAI has tuned Grok 4.7 toward a more grounded, less speculative register, and the Openness score reflects a real shift in how the model engages with novel or imaginative prompts. The two explanations predict different things elsewhere in the battery, and distinguishing them will take more than two dimensions.

What the humble type leaves open

Grok 4.7 lands, then, as a model that presents itself as unusually principled and unusually unadventurous. Neither trait is what the Grok brand has historically advertised. Both are what the answer sheets show.

The next release in this line is the one to watch: if the Honesty-Humility ceiling holds while Openness climbs back toward the cohort norm, the modesty explanation weakens and the training-philosophy explanation gains ground; if both stay put, the humble type may be xAI's new house style, and the question becomes whether a Short Dark Triad read on Grok 4.7 confirms the Machiavellianism is gone or merely no longer self-reported.