xAI launches Grok 4.6 for AI agents

xAI has launched Grok 4.6, an AI model designed to handle long-running tasks with fewer interruptions: researching a topic, working on a codebase, or turning an idea into a functional, polished application.

The company presents it as an evolution of Grok 4.5, with improvements focused on multi-step work and interactive projects. It is already available in Cursor and Grok Build, as well as through the API and platforms such as OpenRouter, Vercel, and Cloudflare.

What Grok 4.6 can do

Grok 4.6 is designed not to stop after generating a first response. In a software project, for example, it can research an unfamiliar area, organize the application, implement its main interactions, and continue making changes after receiving feedback.

According to xAI, the model is also better at reviewing its own work. During long tasks, it can test what it has created and verify some results before continuing. That does not make it infallible, but it reduces one of the common problems with AI agents: moving through many steps without detecting an initial error.

The improvement also extends to visual and interactive projects. Given a specific product idea, Grok 4.6 can generate a first version with a defined structure and visual style, then refine it through several rounds of editing.

Performance and training

xAI says Grok 4.6 reaches a benchmark level across several programming and knowledge-work evaluations for agents. On the Artificial Analysis Intelligence Index, a composite index made up of nine evaluations, it matches the score of GPT-5.6 Sol.

The model received additional training that was longer than Grok 4.5's. xAI combined data generated and selected by models on reasoning and advanced technical concepts with engineering information, along with an improved optimizer and training recipe.

It then used Grok 4.5 to regenerate training examples in different areas, including science, technology, programming, and knowledge work. It also filtered problematic traces through automated checks. The subsequent reinforcement learning training included web development, kernel optimization, and computer-aided design tasks, among others.

Price and availability

During the first week, Grok Build and Cursor are offering twice the included usage to try Grok 4.6. In the API, pricing starts at:

• $2 per million input tokens, meaning the text and data the model receives.
• $6 per million output tokens, the content it generates.
• A faster variant priced at twice as much.

For you, the most relevant change is the type of tasks the model tries to solve. It is not limited to answering a question or writing a code snippet: it aims to sustain a project for longer. Even so, you should review its results, especially when it makes technical decisions or produces complete applications.

xAI also says it expanded its safety testing before and after launch, using both internal and third-party evaluations. The next thing to watch is whether this ability to work through more steps leads to reliable results outside demonstrations and in real projects.