xAI unveils Grok 3: the AI model that can spend minutes thinking to solve complex problems

The artificial intelligence landscape advances with the launch of Grok 3, a model that marks a turning point in AI systems' reasoning capabilities. Developed by xAI, this new model has been trained on the Colossus supercluster, using computing power ten times greater than xAI's predecessor models.

Grok 3's main innovation lies in its ability to dedicate time to reasoning, similar to how a human would. The model can spend from seconds to several minutes analyzing a problem, considering different approaches, and verifying its own solutions before offering an answer. This feature has been achieved through extensive reinforcement training.

The test results demonstrate its exceptional performance. In the AIME 2025 mathematics exam, conducted just a week ago, Grok 3 achieved 93.3% accuracy. Additionally, the model has achieved 84.6% in graduate-level reasoning tests and 79.4% in code generation assessments.

Alongside the main model, xAI has introduced Grok 3 mini, a more efficient version designed for tasks requiring less general knowledge. This variant has achieved outstanding results, reaching 95.8% accuracy in AIME 2024.

The new system will be available to ???? Premium and Premium+ users (formerly Twitter) through the ???? and Grok.com platform. The company has also announced the upcoming release of an API that will allow developers and companies to integrate Grok 3's capabilities into their own applications.

This advance represents a significant step in the development of more sophisticated AI systems, capable of addressing complex problems with unprecedented depth and precision.