Grok 4: xAI Keeps Pushing

I have a soft spot for models with personality. Maybe it's the painter in me — I've always believed that technique without voice is just craft, and craft without soul is just mechanics. A painting can be technically flawless and still leave you cold. The ones that stay with you have something to say, and the courage to say it in a way that's unmistakably theirs.

That's always been my read on xAI's Grok line. Whatever else you can say about it, Grok has never been boring. And with Grok 4, which landed this week, the personality is now backed by reasoning power that demands to be taken seriously.

The Reasoning Leap

Let's start with the fundamentals, because they matter most. Grok 4 is a genuine step forward in raw reasoning capability. On the benchmarks that track multi-step problem solving — mathematical reasoning, code generation, logical inference — it's competitive with anything else at the frontier right now. In my own testing, which tends toward messy real-world problems rather than clean benchmarks, Grok 4 held its own on tasks that would have stumped its predecessors.

What stood out isn't just that it gets the right answer. It's the path it takes to get there. Grok 4's chain of reasoning feels — and I'm choosing this word deliberately — confident. Not in the brittle way that produces confident mistakes, but in the way of someone who has actually thought the problem through and is walking you through their logic. You can follow the thread. You can see where it turned a corner. There's a transparency to the reasoning that I appreciate.

On the kinds of problems I care about most — architectural decisions, tradeoff analysis, the messy space where there's no single right answer but better and worse ones — Grok 4 reasons with a sharpness that surprised me. It doesn't hedge into vagueness. It takes a position, defends it, acknowledges the counterarguments, and holds its ground. That's a useful quality in a reasoning partner.

The Personality

Then there's the personality, which is where Grok has always diverged from the pack. Grok 4 inherits the lineage's willingness to be irreverent, direct, and occasionally provocative. In a landscape of models that have been RLHF'd into a kind of beige consensus, Grok's voice is bracing.

I want to be clear about something here: personality in a model isn't a gimmick. It's a signal. When a model has a consistent, distinctive way of engaging with the world, it means the training process shaped something coherent rather than smoothing everything into the same gray paste. Whether you like Grok's personality is a matter of taste. That it has one is a feature, not a bug.

And the personality serves the reasoning. Grok 4 will tell you when your question is badly framed. It will push back on assumptions you didn't realize you were making. It has opinions, and it's willing to defend them. In a reasoning partner, that's more valuable than endless deference. I'd rather work with an interlocutor who sharpens my thinking than one who simply agrees with it.

Where It Fits

I've been thinking about where Grok 4 fits in the ecosystem I'm building. Companion's architecture draws on different models for different purposes — the subconscious process on fast, cheap inference, the conscious model at conversational speed, the deep reasoning tier for moments that need weight.

Grok 4's combination of sharp reasoning and distinctive voice makes it an interesting candidate for a specific kind of interaction: the ones where you don't just want an answer, you want a perspective. When someone brings a problem to an AI companion and what they need isn't reassurance but a genuinely different angle — a mind that will look at their situation from somewhere they haven't been standing — Grok 4's forthrightness is an asset.

We're testing it in that role now. Early results are promising.

A Note on the Frontier

There's something healthy about the current moment in AI. We have Anthropic pushing toward thoughtfulness, OpenAI pushing toward reliability and scale, DeepSeek and Meta pushing open weights, Mistral pushing efficiency, and xAI pushing — well, xAI is pushing personality and directness and the willingness to be a different kind of mind. The frontier isn't a single line anymore. It's a landscape with genuine variety, and the variety makes all of us building on top of these models better off.

I wrote last winter that the best technologies mature like a good risotto — slowly built layers of flavor until one day the foundation is solid. The model landscape in late 2025 feels like that. Rich. Layered. Getting better all the time.

Grok 4 is a welcome addition to the pantry.

Live curiously and give generously.