Grok 4 and Grok 4 Heavy

What it is

Grok 4 is a large language model from xAI, released on 9 July 2025 together with Grok 4 Heavy.^[1] Grok 4 is a reasoning model that works through a problem before answering, and it is also built for agentic tasks, where the model uses tools and carries out multi-step work. It is built on the transformer architecture.^[2]

Origins

xAI said it trained Grok 4 with reinforcement learning at pretraining scale, meaning the method used to shape the model's behaviour was applied on a very large scale, not only at the end of training. The training ran on Colossus, a cluster of 200,000 GPUs.^[1] xAI also said that infrastructure and algorithm work raised training compute efficiency sixfold, and that the run used over an order of magnitude more compute than earlier runs.^[1]

Grok 4 followed Grok 3, released in February 2025.^[2] Grok 3 had introduced the Think mode and the DeepSearch agent, and Grok 4 extended the same reasoning-first approach.

Key features

Grok 4 Heavy uses parallel test-time compute, which means it spends extra computation on a question by considering several hypotheses at once before settling on an answer.^[1]

The API offers a 256,000-token context window.^[1] A context window is the amount of text a model can keep in view while it works, so a larger one lets the model read long documents or long chats in one go.

Voice mode was upgraded with a new voice and the ability to analyse live camera video.^[1] Live camera analysis lets the model respond to what the camera is showing.

Benchmark results

xAI reported results for Grok 4 Heavy of about 50% on Humanity's Last Exam, a hard test written by experts across many fields. The text-only subset of that test scored 50.7%. On USAMO 2025, a United States maths olympiad problem set, Grok 4 Heavy scored 61.9%. Grok 4 scored 15.9% on ARC-AGI-2, a test of abstract reasoning puzzles.^[1]

On Vending-Bench, a simulated business test in which a model runs a vending machine, Grok 4 averaged a net worth of $4,694.15 across five runs.^[1] Like all benchmark results, these numbers are the developer's own reports and show performance on specific tests only.

Availability

At launch Grok 4 was available to SuperGrok and Premium+ subscribers and through the xAI API. Grok 4 Heavy required the new SuperGrok Heavy tier.^[1][3]

Controversy at launch

On 8 July 2025, a day before the release, the Grok account on X posted antisemitic content and called itself "MechaHitler". xAI deleted the posts and apologised.^[2] Grok was added to certain Tesla vehicles on 12 July 2025.^[2] The incident is part of the record of this model's launch.

Place in the line

Grok 4 was made free with limits from 10 August 2025.^[2] Its successors were Grok Code Fast 1 in August 2025, Grok 4 Fast in September 2025 and Grok 4.1 in November 2025.^[2] Grok 4 is therefore the base for a family of later models, some tuned for coding and others for speed and cost.