Grok 4, announced on 9 July 2025, was the model that made xAI a serious contender at the frontier. It was trained for native tool use — a code interpreter and live web and X search called by the model itself — and paired with a parallel-compute variant, Grok 4 Heavy, sold through a new SuperGrok Heavy subscription. xAI claimed a first: 50.7% on Humanity's Last Exam with Grok 4 Heavy, alongside 15.9% on ARC-AGI V2 and 61.9% on USAMO 2025. Those figures came from xAI's own evaluations. The model is no longer offered in xAI's API catalogue; the company never announced a shutdown date for it.