xAI, the artificial intelligence company founded by Elon Musk, has officially launched its latest flagship models, Grok 4 and Grok 4 Heavy. Alongside the release, the company introduced its most premium subscription tier to date, SuperGrok Heavy, which is priced at $300 per month and targeted at power users and developers.
Grok 4 is designed to compete with top-tier AI models like OpenAI’s ChatGPT and Google’s Gemini. As per xAI, Grok 4 demonstrates frontier-level performance across several benchmark evaluations.
On Humanity’s Last Exam, an extensive test designed to evaluate general knowledge, Grok 4 scored 25.4% without using external tools, outperforming OpenAI’s o3 (high), which scored 21%, and Gemini 2.5 Pro, which scored 21.6%.
When equipped with tools, Grok 4 Heavy scored 44.4%, making it the top-performing model in that category. For context, Gemini 2.5 Pro with tools scored 26.9%, as noted in a TechCrunch report.
xAI also highlighted Grok 4 Heavy’s “multi-agent” system, which enables multiple instances of the model to solve a problem in parallel and compare answers, similar to a study group. This collaborative approach is intended to improve both reasoning and accuracy in complex tasks.
The company also shared results from the ARC-AGI-2 benchmark, where Grok (Thinking) achieved new state of the art (SOTA) score on the test with 15.9%. As per Arc Prize in a post on X, this score is “nearly double(s) the previous commercial SOTA and tops the current Kaggle competition SOTA.”
The $300/month SuperGrok Heavy plan grants users early access to Grok 4 Heavy, as well as upcoming product features. Upcoming capabilities that are planned for release in the coming months include an AI coding model in August, a multi-modal agent in September, and a video generation system in October.
Grok 4 and Grok 4 Heavy are available via API as part of xAI’s push to engage developers and enterprise users. The company’s enterprise platform, which launched just two months ago, will also expand via partnerships with cloud hyperscalers to bring Grok models to broader infrastructure environments.