Grok 4 is positioned by xAI as their most intelligent model, extending the trajectory from Grok the listed price by scaling reinforcement learning beyond its earlier reasoning variant. Training was carried out on Colossus, a roughly 200,000-GPU cluster, with compute efficiency improvements of about 6x and an expansion of verifiable training data from primarily math and coding into many more domains. xAI describes the training run as consuming over an order of magnitude more compute than prior efforts, producing smooth performance gains as reinforcement learning was applied at pretraining scale.

The release emphasizes native tool use and real-time search integration as defining capabilities, with availability through SuperGrok and Premium+ consumer subscriptions and through the xAI API. A higher-end SuperGrok Heavy tier exposes Grok 4 Heavy as the most powerful configuration in the family, and the announcement page also previews Grok 4 Voice Mode as part of the roadmap. In practice, this makes Grok 4 well suited to tasks that combine reasoning with live information retrieval and tool-driven workflows, where integration depth matters as much as raw generation quality.