From the source
xAI announces Grok 4, described as the most intelligent model with native tool use and real-time search integration.
It is available to SuperGrok and Premium+ subscribers and through the xAI API.
A new SuperGrok Heavy tier provides access to Grok 4 Heavy, which employs parallel test-time compute.
Grok 4 was trained using reinforcement learning on the Colossus cluster (200,000 GPUs) with a 6x improvement in compute efficiency.
Grok 4 Heavy is the first model to score 50% on Humanity's Last Exam and achieves state-of-the-art results on benchmarks including ARC-AGI V2 (15.9%), Vending-Bench ($4694.15 net worth), and USAMO'25 (61.9%).
