xAI’s Grok 4 announcement treats tool use as part of the model’s reasoning loop rather than a separate product attachment. The company says Grok 4 was trained with reinforcement learning to use a code interpreter and web browsing, and that it can choose search queries while working through difficult questions. That design raises the bar for a model evaluation: a system may be judged not only by the answer it gives, but by how it decides to gather and check information.
The release describes reinforcement learning at pretraining scale, supported by xAI’s 200,000-GPU Colossus cluster. It also reports a sixfold increase in compute efficiency from infrastructure and algorithmic work. These are descriptions from xAI, not an independently verified account of training. They nevertheless show where frontier labs believe advantage is moving. More compute is being paired with verifiable rewards and tool-mediated tasks, extending the training target beyond next-token prediction.
xAI reports that Grok 4 reached 15.9% on ARC-AGI V2 and that Grok 4 Heavy scored 50.7% on Humanity’s Last Exam. The company also reports results for an agentic business simulation called Vending-Bench. Provider benchmarks can reveal what a lab wants to optimize, but they do not settle whether a model will behave reliably in unfamiliar settings. Tool access can improve factuality when the search is well chosen, while a poor query or stale source can make an answer look confident without making it sound.
The release also pairs capability with product reach. xAI says Grok 4 is available to subscribers and through its API, with a 256,000-token context window described for the API. This article does not treat those interfaces as a tutorial; the relevant point is that a model with live information access changes the timing of an answer. Its knowledge is no longer fixed at training time, so evaluation must include provenance, retrieval decisions, and how the system signals uncertainty.
Grok 4’s significance is therefore methodological as much as competitive. A tool-using model can solve a harder class of problems, but its chain of actions becomes part of the evidence readers need. Future comparisons will need to report which tools were available, what the model chose to consult, and whether independent reviewers can reproduce the result. The frontier is moving from isolated responses toward observable reasoning under changing information.