The SpaceXAI team has officially unveiled Grok 4.6, a new flagship model that focuses on performing long-term agency tasks, writing software code, and creating interactive applications. This version is based on Grok 4.5, released in early July, but the developers have shifted the focus from generating individual responses to the neural network's ability to work long and consistently on one complex project. The model is now able to independently explore an unfamiliar subject area, design the architecture of the application, implement its main functions, test the result and repeatedly refine it based on feedback. Company representatives emphasize that when performing long chains of actions, the new model is much more likely to independently verify its work before moving on to the next stage.

According to the test results, Grok 4.6 scored sixty-one points in the reputable Artificial Analysis Intelligence index, equaling the GPT-5.6 Sol model in this complex indicator. Both of these neural networks are now second only to Fable 5 Max in first place. To achieve these results, the SpaceXAI team has significantly increased the duration of the additional training stage compared to the previous version. This process used carefully selected data from the model itself, concerning complex logical reasoning and advanced technical concepts. One of the main directions of development was the creation of full-fledged applications based on the general description of the idea, when the neural network itself determines the structure of the project and implements all the necessary interactions. In visual and interactive projects, the new version demonstrates a much stronger initial result, immediately setting a competent structure and visual language of the application before starting iterative improvements.

In addition, Grok 4.6 has received much broader training for agent-based work, especially in those scenarios where artificial intelligence does not just respond to a text query, but independently performs complex chains of actions. The new model is already available to developers through the API, Grok Build platform, Cursor editor, and a number of other partner services. The basic cost of use starts at two dollars per million input tokens and six dollars per million output tokens, while the accelerated version will cost customers twice as much. This release was a logical continuation of the company's recent announcements, because just the day before, the SpaceXAI team introduced Grok Bot, a specialized AI agent system for performing complex work tasks in a convenient messenger format. The technology giant continues to actively develop the ecosystem of autonomous assistants, blurring the boundaries between text generation and the actual execution of applied tasks.