Grok 4
xAI's flagship reasoning model with native tool use and live search
Grok 4 is a large language model from xAI, released in July 2025 as the successor to Grok 3. The company says it was trained with reinforcement learning at pretraining scale on a 200,000-GPU cluster, with native tool use and real-time search across the web and X.
- Trained with reinforcement learning on Colossus, a 200,000 GPU cluster
- xAI reports 50.7% on Humanity's Last Exam text-only subset
- 256,000-token context window with text and vision input
- Grok 4 Heavy runs several hypotheses in parallel at test time
- Native tool use: code interpreter, web browsing and search inside X
What it is
Grok 4 is a general-purpose language model built by xAI, the company founded by Elon Musk. It arrived in July 2025 as the next step after Grok 3 and Grok 3 Reasoning. xAI describes it as its most intelligent model, aimed at hard reasoning tasks, research questions and agentic work. The model writes its own search queries, calls a code interpreter and browses the web while it thinks. A stronger variant, Grok 4 Heavy, runs several hypotheses in parallel. The company positions it for both consumers, through SuperGrok and Premium+ subscriptions, and developers, through the xAI API.
How it came about
xAI's earlier work scaled next-token prediction pretraining with Grok 3, then added Grok 3 Reasoning, which used reinforcement learning to think longer before answering. While developing that reasoning model, the team saw scaling trends suggesting reinforcement learning could be pushed much further. Grok 4 was the result: training ran on Colossus, xAI's 200,000-GPU cluster, with more than an order of magnitude more compute than previous runs. The company says stack-wide improvements raised training compute efficiency sixfold, and that it greatly widened its set of verifiable training data beyond maths and coding.
Inside the model
Grok 4 was trained with reinforcement learning applied at pretraining scale, refining reasoning rather than only predicting the next token. xAI expanded verifiable training data, where answers can be checked automatically, from mainly maths and coding into many more domains. The model was also trained to use tools: a code interpreter, web browsing, and search inside X using keyword and semantic queries, plus viewing media there. Grok 4 Heavy adds parallel test-time compute, weighing several hypotheses at once. The API offers a 256,000-token context window and multimodal understanding across text and vision.
What it can do
Grok 4 scores 15.9% on ARC-AGI V2, which the company calls a new high for closed models. On the agentic Vending-Bench, across five runs, it reports $4694.15 net worth and 4569 units sold, against $2077.41 and 1412 units for Claude Opus 4 and $844.05 and 344 units for humans. These are the company's own figures.
How to use it
Grok 4 is sold to SuperGrok and Premium+ subscribers and offered through the xAI API. Grok 4 Heavy sits in a separate SuperGrok Heavy tier. The API includes live search across X, the web and news sources, and xAI says it holds SOC 2 Type 2, GDPR and CCPA compliance. The company also says Grok 4 is coming to hyperscaler partners. No open weights are offered.
Safety and testing
xAI states that its API carries enterprise-grade security and compliance, naming SOC 2 Type 2, GDPR and CCPA certifications, and describes these as protection for sensitive applications. The announcement does not describe safety evaluations or red-teaming for Grok 4 itself.
Reception and debate
The launch announcement makes strong claims about benchmark leadership, including being the first model to score 50% on Humanity's Last Exam and setting a new high for closed models on ARC-AGI V2. These are xAI's own reported results rather than independent assessments, and no third-party reception is described here.
Why it matters
xAI says Grok 4 continues a trend of scaling reinforcement learning far beyond earlier runs, and plans to push further, moving from rewards that can be verified in controlled settings toward messy real-world problems. It also intends to broaden multimodal work, covering vision and audio, while making models faster and more efficient.
Questions people ask about Grok 4
What is Grok 4?
It is xAI's flagship language model, released in July 2025 after Grok 3. It reasons through problems with reinforcement learning training, writes its own web searches and calls tools such as a code interpreter while answering.
How is Grok 4 different from Grok 4 Heavy?
Grok 4 Heavy is the more powerful version. It uses parallel test-time compute, meaning it weighs several hypotheses at the same time before answering. It is sold in a separate SuperGrok Heavy subscription tier.
How do I get access to it?
Consumers can use it through SuperGrok or X Premium+ subscriptions. Developers can call it through the xAI API, which includes live search across X, the web and news sources. xAI says it is also coming to hyperscaler partners.
What can it take as input?
Through the API it handles text and vision with a 256,000-token context window. In the app, an upgraded Voice Mode adds speech, and a camera feature lets Grok comment on what it sees during a voice chat.
How large is it?
No parameter count is given. xAI only says training used its 200,000-GPU Colossus cluster, with over an order of magnitude more compute than earlier runs and a sixfold gain in training compute efficiency.
Good reading and sources
There is no English Wikipedia article on Grok 4 yet, so the facts on this page come from the sources listed above: x.ai/news/grok-4. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.