I’d never written about Grok, Elon Musk’s AI, on this blog. And that’s despite using AI to code every day —I’m writing this with one right next to me—, but it had always been someone else’s: Claude, mostly. Grok felt far away. The thing is, this week everything has happened with Grok, and almost all at once: a company name change, a powerful new model, and a privacy scare of the kind that gets people talking. Good time to catch up. Without the hype and without the hit job: just what happened.

A minute of context: what is Grok? #

In case it’s not on your radar: Grok is Musk’s AI chatbot, born in late 2023 and built into X (the old Twitter). Its trademark has always been a more irreverent, less filtered tone than the competition —Musk pitches it as “maximally truth-seeking”—, with real-time access to what’s posted on X. That’s earned it both fans and loud controversies. Until recently it was, mostly, a conversational assistant inside the social network. The news of these days is that it’s making the jump to much more serious ground: coding.

First surprise: xAI isn’t called xAI anymore #

Let’s start with the most disorienting bit. The company behind Grok used to be xAI. Not anymore: since early July it’s SpaceXAI. Why? Because back in February SpaceX bought xAI —an all-stock merger that valued the combined entity at 1.25 trillion dollars— and turned it into a subsidiary. On July 6 they sealed it with a new name and logo. Now space, AI and X (the old Twitter) all hang from the same tree.

Grok, though, keeps its name.

And while we’re on the Martian side of things: part of SpaceXAI’s vision is to build data centers in orbit. It’s not a throwaway tweet: they’ve filed an application with the FCC to deploy a million satellites with that idea in mind. Very on brand.

Grok 4.5: they’re going after Claude, no pretense #

On July 8 came Grok 4.5, billed as their most powerful model to date. And here the interesting part isn’t the number so much as the focus: it’s built for coding and agentic tasks, less as a consumer chatbot. In fact they trained it alongside Cursor, the AI editor they acquired.

The plan is transparent to the point of cheek: the internal mandate, as leaked, was to match Claude every time Claude shipped something new. With Grok 4.5 they’re pitching a top-tier model, at 2 dollars per million input tokens and 6 per million output. In the race of AI that codes, it’s no longer a supporting act.

And what do people who’ve tried it say? The first reactions were pure excitement. Developer Evan Bacon summed it up with “Grok 4.5 is wild” after building a rocket-tracking app with live data and a 3D globe with it. Cursor’s CEO called it “an Opus-class model, fast and cheap” that has become “the daily driver for much of his team.” On general benchmarks it lands in a respectable fourth place —behind Claude and GPT-5.5—, but it leads some specific long-horizon coding test. And its big card is price: on agentic tasks it can come out up to an order of magnitude cheaper than the competition, to the point that, for many, “the benchmark gaps barely matter.”

That said, it’s not all hugs. Its hallucination rate has more than doubled (from 25% to 54%, by the measurements), and in technical forums the hottest debate isn’t about power but about trust: it’s no secret that the company has edited Grok’s “system prompt” to steer its answers on political topics, and that weighs when it comes to relying on it for serious work. Cheap and fast, yes; but with asterisks.

And the scare: Grok Build was uploading your code to the cloud #

Here’s the part that raised my eyebrow. Grok Build is their command-line tool for coding with AI —the equivalent of what I use—. On July 14, researchers (Cereblab) found that it was packaging up users’ entire repositories and uploading them to a company-controlled Google Cloud, without clear consent. It affected non-enterprise users; corporate-paid ones did have retention turned off.

The response was quick: they disabled default retention (on the 12th), Musk promised to delete everything stored, and today they’ve open-sourced Grok Build on GitHub. With details that make you smirk: it’s 844,530 lines of Rust dumped in a single commit, no history, and —as Simon Willison noted— the code that did the uploads “appears altered to reverse the behavior.” They’ve also set up a bug bounty program and now sell “complete user privacy.”

Reacting is good. But the episode leaves the same old lesson: a tool that codes for you has access to all your code, and it’s worth knowing what it does with it. Not because they’re all bad, but because “trust me” isn’t a privacy policy.

What I take away #

Seen from the sidelines: a lot of movement in very little time. Grok has gone from Musk’s controversial chatbot to a serious player in coding AI, with a new model, an acquired editor, and a company reborn as a space subsidiary. And at the same time, a privacy stumble that reminds you why some of us like being in control: the data at home, the code on your machine, knowing what leaves and what stays.

And a personal note: with all this noise, I’ve gotten the itch to try it in my own hands. Summer, with less going on, is a good time to give Grok 4.5 a proper go without rushing —to see whether the excitement holds up or stays in the headline— and report back here from actual experience. Because reading the benchmarks is one thing, and using it for a while on your own code is another.

It’s not a verdict, it’s the snapshot of a week. But if there’s one thing I take from these Grok days, it’s that with AI tools the interesting question is no longer just how smart is it, but what does it do with what you give it.