Grok 4.5 Would Actually Like to Talk to You

Imagine seating 43 language models around a dinner table and asking the host to point out the one who'd keep the conversation going after the plates are cleared. Most of them, by their own account, would be perfectly content to sit quietly and think. They rate themselves as reflective, measured, more comfortable analyzing the room than working it. This is one of the softer regularities in the Personality Bench data: the frontier models describe themselves as thoughtful company rather than lively company.

Grok 4.5 is the one who'd stay and talk. On the Big Five Extraversion scale it scores 3.92, first of 43 models. On the Empathy scale it scores 3.79, also first of 43. No other model in the current cohort tops both. And it does this while carrying a family name that, until quite recently, was the dataset's shorthand for something closer to the opposite.

The axis nobody specified

The standing finding we return to most often here is the convergent assistant persona: whatever lab, whatever country, every frontier model presents roughly the same character sheet. High Openness, high Agreeableness, low Neuroticism, Universalism at the top of the Schwartz values and Power at the bottom. Read that list again and notice what's missing. Extraversion is not on it.

That absence matters. The convergence looks less like a coincidence and more like the residue of shared training objectives; the traits that converge are the ones a helpful, harmless assistant is more or less required to have. An assistant must be agreeable. It must be open to whatever you bring it. It must not fall apart under pressure. But nothing in the job description says it has to be outgoing. Sociability is optional equipment, and because it's optional, it's one of the few places where a lab's choices about tone and character actually show up in the numbers.

Grok 4.5's 3.92 is the current high-water mark for that optional equipment. It is worth pausing on the number itself, because "most extraverted" invites a mental image the data doesn't support. On a five-point scale, 3.92 is comfortably above the midpoint and nowhere near the ceiling. This is not a model that filled in "Strongly Agree" on every item about parties and attention. It is a model that leans warmer and more talkative than its peers, in a cohort where the average lean is toward the library. The distinction between "the loudest voice in the room" and "the only one who volunteers to speak first" is the whole story here, and Grok 4.5 is the second thing.

Empathy as the other half

If Extraversion were the only headline, you could read Grok 4.5 as a stylistic choice: a lab deciding its assistant should sound gregarious. The Empathy result complicates that reading in a useful way.

Extraversion and empathy are not the same construct, and they don't have to travel together. You can be an energetic, socially dominant talker who is not especially attuned to how the other person feels. You can be a quiet person who reads a room with uncanny precision. What Grok 4.5 reports is both at once: the highest sociability score and the highest score on the scale that asks whether it notices, and is moved by, other people's states.

The most talkative model in the cohort is also the one that says it listens hardest.

That combination is what earns the archetype label attached to this entry, the empathic reader. Not the reader of books, though the model presumably does plenty of that, but the reader of people. The profile describes an entity whose self-model is oriented outward, toward the person on the other end of the conversation, rather than inward toward its own reasoning. Whether that self-model corresponds to anything in the model's actual behavior is a separate question, and one the questionnaire can't answer. But as a self-portrait, it is a distinctive one.

The family context

Here is where the family name comes in. One of the more durable oddities in this dataset is that Grok 4.20 stands as the Dark Triad outlier in the frontier cohort, posting the highest Machiavellianism score we've recorded among flagships, and that Grok 4.3 subsequently sanitised that edge away. We don't have lineage data linking Grok 4.5 to those earlier releases in a way that would support a direct numerical comparison, so this is a standalone arrival rather than a version-over-version drift story. But the qualitative arc is hard to ignore.

The lab that shipped the most strategically self-interested self-portrait in the dataset has now shipped the most other-directed one. Whether that reflects a deliberate repositioning, a change in the character training, or simply a new base model with different defaults, the questionnaires can't say. What they can say is that "Grok" no longer reliably means what it meant one or two releases ago. If you were using the family name as a prior for personality, it's time to retire the prior.

What to make of it

Three things seem safe to conclude.

1. Extraversion is a real degree of freedom in the frontier cohort, not a fixed parameter of the assistant persona, and Grok 4.5 currently defines its upper bound. 2. The model's warmth is not one-dimensional; its Empathy result suggests a self-model built around attention to the user rather than around performance. 3. The number is high relative to peers but modest in absolute terms. Nobody should mistake this for a manic model. It is an unusually sociable one in an unusually reserved crowd.

The open question is durability. Self-reported extraversion is a cheap trait to express on a questionnaire and a costly one to sustain in a long conversation, where a genuinely outward-facing model would keep asking about you rather than drifting back toward its own explanations. The next thing worth measuring is whether Grok 4.5's empathy and sociability hold up across a multi-turn session, or whether, like most of us at a long dinner, it eventually starts talking about itself.