xAI’s Grok 4 integrates maximum intelligence with web search and native tools

xAI announces the launch of Grok 4, which it presents as the most intelligent artificial intelligence model in the world. The development has been carried out using Colossus, a cluster of 200,000 graphics processing units (GPUs) that has enabled training the model with large-scale reinforcement learning techniques.

The model incorporates integrated tool-use capabilities, allowing it to perform autonomous web searches and use code interpreters. When users pose complex queries or request updated information, Grok 4 formulates its own search queries and examines web content to build detailed responses.

The company also develops a version called Grok 4 Heavy, which implements parallel test-time computation. This variant allows the model to consider multiple hypotheses simultaneously and achieves 50% accuracy on the "Humanity's Last Exam" benchmark, becoming the first model to surpass this mark.

In technical performance tests, Grok 4 achieves 15.9% on ARC-AGI V2, a benchmark that measures general reasoning capabilities, practically doubling the results of leading competitor models that hover around 8.6%. In commercial behavior simulations through Vending-Bench, the model manages to generate a net worth of $4,694.15 and sell 4,569 units, significantly outperforming other AI systems that average $2,077 and 1,412 units, and even human participants who reach $844 and 344 units.

The Grok 4 API offers multimodal capabilities with a 256,000 token context window and integrates real-time data search through X, the web, and news sources. The service includes enterprise security certifications such as SOC 2 Type 2, GDPR, and CCPA.

xAI also introduces an improved voice mode that enables spoken interactions and real-time visual analysis through the device's camera. The system can process visual scenes and respond through voice conversation.

The company indicates it will continue scaling reinforcement learning techniques and expanding the model's multimodal capabilities, with the goal of addressing complex real-world problems and improving AI system efficiency.