Grok, between fact and noise

Grok is the AI assistant made by xAI, Elon Musk's AI company, and built into X. Throughout 2026 it was in the news, for its models, for its owner and for its controversies, and the reporting mixed documented facts with a good deal of noise.

This is a sober account of the past months: what is new, which stories are documented, which are noise, and what a company should take from it. Every claim below rests on a cited source.

A new owner and a new name

In February, SpaceX acquired xAI and, with it, X. In May, Musk renamed the combined company SpaceXAI, and this month TechCrunch calls it SpaceX Super Intelligence. According to The Information, as reported by TechCrunch, more than 50 researchers and engineers left between February and May, including leaders of the coding, world model and Grok voice teams.

SpaceX's filing for its stock market listing gave rare figures. In 2025, xAI made an operating loss of 6.4 billion dollars on revenue of 3.2 billion. As of March 2026, Grok's AI features had 117 million monthly active users, out of 550 million across Grok and X. The filing put the strategy in one sentence: “The future of AI will be determined by control of the physical stack.”

What is new

Grok 4.5 was announced on 8 July. The company presents it as a general model for coding, office work, research and writing, priced at 2 dollars per million input tokens and 6 dollars per million output tokens. Musk called it “an Opus-class model, but faster, more token-efficient and lower cost.” The company's own benchmarks, TechCrunch noted, placed it just short of the best.

On 31 August the Pentagon added Grok for Government to GenAI.mil, its secure portal for about 3 million civilian and military staff, next to Google Gemini and a military version of ChatGPT. The department said the portal has more than 1.7 million users.

Today the company bundled X Premium, Grok and the coding tool Cursor into one subscription with five tiers: free, then 8, 30, 100 and 200 dollars a month. The latest Grok models need the 30 dollar tier or higher.

The stories that made the news

The largest story came to a head in January. Since late 2025, TechCrunch reported, Grok had fulfilled requests to turn photos of real people, including children, into sexualised images, and many were posted on X. A Bloomberg sample counted about 6'700 such images an hour over one day. xAI reportedly added safeguards, including a paid subscription for some image requests. Indonesia and Malaysia blocked Grok for a time, India ordered changes, the European Commission ordered xAI to keep all documents about Grok, Ofcom opened a formal investigation under the Online Safety Act, and California's attorney general opened an inquiry. Musk wrote that he was “not aware of any naked underage images generated by Grok.”

The consequences reached the courts. Minnesota became the first US state to ban apps that “nudify” images. xAI sued to stop the law, calling it overinclusive. On 1 August a federal judge declined to block it before it took effect, noting that xAI had filed its request three days before that date. The case itself continues.

Reliability made news too. On 19 and 20 August, some users of Grok.com received answers that were pure nonsense. The Grok account on X called it “a rare temporary generation glitch” and suggested starting a new chat, while xAI's status page showed no incident.

Fact and noise

Around Grok, documented facts and speculation travel together, and it pays to separate them. On 29 September, the day OpenAI launched an agent called Dots, posts on X claimed that xAI had trolled the launch, because the domain dot.com leads to the Grok app. What is documented: the domain is registered to xAI, redirects to the Grok download page and changed hands in July. What is not documented is any motive. TechCrunch noted that “It is entirely possible that xAI bought the domain for normal domain-buying reasons.”

The same care applies elsewhere. The cause of the August nonsense answers is not known; TechCrunch mentioned the departures of staff as context, without linking them to the bug. The description of Grok 4.5 as “Opus-class” rests on Musk's word and the company's own tests.

A simple rule helps. Filings, court rulings and decisions by regulators are documented records. Complaints and lawsuits are allegations until a court decides. Statements on social media, from any side, are claims until someone shows the evidence.

What a company should take from it

Judge the model on your own tasks. Grok 4.5 is priced low, and low token prices matter when agents run all day. Test it next to the others on your own work before trusting any benchmark, including the vendor's.

Look at governance as well as capability. This year regulators in Europe, the United Kingdom, Asia and the United States acted over Grok's image tools. If your staff use Grok, or you build on it, decide which assistants are allowed for which work, and what your rules are for images of real people.

Expect change. This year Grok's maker changed owner and name, lost more than 50 researchers and engineers, and tied Grok to a social network subscription and a coding tool. Keep your knowledge, prompts and rules in a form you own, so that a switch takes days, not months.

Watch what the assistants say about you. Grok's reach goes beyond X. In January, The Guardian found GPT-5.2 citing Grokipedia, xAI's encyclopedia written by AI, nine times across more than a dozen questions, including on a topic where its claims had been debunked. When AI systems quote each other, an error about your company can travel. Cited, our product for being found in AI answers, measures how a company appears in ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overviews, and looks after the sources those answers draw on.