xAI launches Grok 4 with real-time web search

xAI has launched Grok 4, an artificial intelligence model that can search the internet and X while responding, use tools such as a code interpreter, and analyze images in real time. It is already available to SuperGrok and Premium+ subscribers, and through the xAI API.

The company presents Grok 4 as its most capable model to date. Its main difference is not just its ability to generate text, but that it has been trained to decide when it needs to consult external sources, run code, or search for additional information before answering.

A model trained to reason for longer

xAI used its Colossus supercomputer, made up of 200,000 GPUs, to train Grok 4 through reinforcement learning. This technique rewards correct answers and adjusts the model's behavior during training.

According to the company, improvements to its infrastructure and algorithms multiplied the efficiency of its computations by six. It also expanded the verifiable data used for training. Previously, this data focused mainly on mathematics and programming. It now covers more areas.

The result is a model designed for problems that require several reasoning steps. Instead of answering immediately, it can propose different hypotheses, check them with tools, and continue searching until it has gathered enough information.

Grok 4 Heavy aims for greater reliability

Alongside the main model, xAI is releasing Grok 4 Heavy, a version that uses more computing capacity while generating a response. It can analyze several possibilities in parallel before choosing a solution, a process known as additional test-time compute.

xAI says Grok 4 Heavy reached 50.7% on Humanity's Last Exam, an academic test designed to measure difficult knowledge across different disciplines. It also scored 61.9% on USAMO'25, a mathematics competition.

In other evaluations announced by xAI, Grok 4 scored 15.9% on ARC-AGI V2, compared with approximately 8.6% for Claude Opus 4. In Vending-Bench, a business simulation, it achieved an average net worth of $4,694.15 across five runs, compared with $2,077.41 for Claude Opus 4 and $844.05 for human participants.

These results were reported by xAI itself and depend on the specific conditions of each test, such as access to Python, the internet, or additional tools. They do not automatically translate into better performance on every everyday task.

What you can do with Grok 4

The user version includes several practical features:

• Search for up-to-date information on the web, X, and other news sources.
• Query content on X using keyword and meaning-based searches.
• Run code to analyze data or solve problems.
• Understand text and images in the same conversation.
• Use your phone's camera to identify objects, scenes, or documents while you talk to Grok.

The API offers a context window of 256,000 tokens, meaning it can work with large amounts of information in a single session. It also includes real-time search and multimodal capabilities, allowing an application to analyze text and images and consult recent data.

For businesses, xAI says the API has certifications and compliance requirements such as SOC 2 Type 2, GDPR, and CCPA. The company also plans to bring Grok 4 to cloud infrastructure providers to facilitate deployment at a larger scale.

A new way to use the assistant

Grok is also introducing a redesigned voice mode, with faster responses and a new voice. The camera turns the conversation into something closer to visual assistance: you can point it at an object, screen, or place and ask it to interpret what it sees immediately.

For you, the most important change is that Grok 4 no longer relies only on what it learned during training. It can search for information, check facts, and use tools. That makes it more useful for news, research, and technical tasks, although you still need to review its sources and answers when accuracy matters.

xAI plans to continue expanding reinforcement learning toward real-world problems and improving its vision, audio, and other interaction capabilities. The next challenge will be proving that these improvements in controlled tests also hold up when Grok works with incomplete information, conflicting sources, and real-world tasks.