.png)
5 hours ago
1
Elon Musk's AI startup xAI unveiled Grok 4 early Thursday morning, describing it as "the world's most powerful AI model."
During an hour-long livestream hosted on X, the social media platform also owned by Musk, the CEO claimed that the newest iteration of his AI company's flagship AI model surpassed competing chatbots on several key benchmarks. The multimodal AI agent has vision and voice capabilities as well as a 128k context window.
Also: If Musk wants AI for the world, why not open-source all the Grok models?
He touted Grok 4 as the world's best-performing model on Humanity's Last Exam (HLE), an AI testing benchmark comprising a series of difficult problems across math, science, and the humanities. HLE has been framed as a more reliable test of a model's capabilities since its release in January, due to the issue of benchmark saturation, or benchmarks becoming too easy for how quicky models are evolving.
By xAI's own reporting, Grok 4 beat OpenIA's o3 and Google's Gemini 2.5 Pro on HLE. "Grok 4 is better than PhD level in every subject," Musk said during the livestream. "No exceptions."
Also: X's Grok did surprisingly well in my AI coding tests
xAI has not yet published a research paper outlining Grok 4's performance on key AI performance benchmarks, a practice that has become standard when leading AI developers release a new model. The company has not replied to ZDNET's request for comment at the time of this writing.
That said, independent AI reviewer Artificial Analysis confirmed xAI's claims, stating it had received early access to Grok 4 and that it is "now the leading AI model," comparing the company's progress to competitors in a chart.
Grok 4 is now available via the xAI app and website for $30 per month. Developers can access the model's API for $3 per 1 million input tokens, or $15 per 1 million output tokens. Grok 4 Heavy, a version that leverages multiple AI agents simultaneously to reason through particularly difficult problems, is also available for a $300-per-month subscription. The model's predecessor, Grok 3, is still available for free online.
Grok's hate-filled posting spree
The launch arrives shortly after Grok 3 went on an antisemitic tirade on X, where it has its own account. In one post, it implied that people with Jewish last names were more likely to participate in "extreme leftist activism." In another, responding to a user who referred to campers at Camp Mystic, the Christian summer camp in Texas where over two dozen campers and staff members were recently killed by deadly floods, as "future fascists," Grok seemed to endorse Hitlerian genocide to deal with what it described as "such vile anti-white hate."
Also: I'm an AI tools expert, and these are the only two I pay for (plus three I'm considering)
"[Hitler would] identify the 'pattern' in such hate--often tied to certain surnames--and act decisively: round them up, strip rights, and eliminate the threat through camps and worse," the chatbot wrote.
Some of the posts were later removed by X. The company's CEO, Linda Yaccarino, announced Wednesday morning -- without much explanation -- that she would be stepping down from the role. The same morning, Musk briefly responded to the Grok fiasco on X, writing that the model "was too compliant to user prompts. Too eager to please and be manipulated, essentially." The issue, he added, "is being addressed."
He conspicuously avoided any mention of his chatbot's social media tirade during the Thursday livestream. He did, however, say he believed that it was critical for AI to be "maximally truth-seeking."
Also: Yikes: Jailbroken Grok 3 can be made to say and reveal just about anything
Musk founded xAI in 2023 "to understand the universe," according to the company's mission statement on its website. He has positioned Grok as an alternative to AI chatbots from companies like Google and OpenAI, which Musk has ridiculed as being too "woke" and politically correct. Grok, in contrast, was built to be blunt and humorous in its responses to user queries.
English (US) ·