xAI launches Grok 4.6 for smarter AI agents and coding

xAI has introduced Grok 4.6, its latest model designed to handle longer and more complex tasks with greater reliability. The release builds on Grok 4.5 and focuses on areas including autonomous agents, coding, visual projects and knowledge-based work.

The central idea behind Grok 4.6 is persistence. Instead of simply responding to one prompt, the model is designed to work through multiple stages of a task, whether that means researching a topic, analysing information, navigating a codebase or turning an early idea into a working application.

What makes Grok 4.6 different

According to xAI, Grok 4.6 has been trained to perform better on agentic tasks. Agentic AI refers to systems that can plan actions, use tools, check their progress and continue working towards a goal with less continuous human direction.

That capability could be particularly useful for developers, researchers and businesses handling multi-stage workflows. xAI says Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite evaluation based on nine benchmarks.

The company also reports improvements over Grok 4.5 on evaluations including CursorBench, DeepSWE, FrontierCode and APEX-Agents. Benchmark results, however, should be viewed alongside real-world performance, as different tests measure different capabilities.

More capable coding and product building

Software development is one of the major areas targeted by the new model. xAI says Grok 4.6 can work across larger codebases, understand broader project objectives and help create functional first versions of applications.

It is also designed to perform more self-testing and verification during extended tasks, allowing it to check parts of its work before continuing. The model is also aimed at visual and interactive projects. Given a product concept, it can help structure an application, establish its visual direction and implement core interactions.

This could make Grok 4.6 useful for rapid prototyping, where teams want to move from an idea to a working first version before spending more time on refinement.

Longer training and wider availability

As part of its development, Grok 4.6 underwent an extended supplemental training phase that went beyond the training run for Grok 4.5. The process included reasoning data, technical concepts, engineering material and reinforcement learning tasks covering areas such as coding, web development, computer-aided design and kernel optimisation.

The AI firm also says it has updated the model's safeguards to account for its expanded capabilities. Grok 4.6 is accessible through Grok Build, Cursor, and the xAI API, as well as through partners including Cloudflare, OpenRouter, and Vercel.

Usage is priced at $2 per million and $6 per million for input and output tokens, respectively. Moreover, the faster variant carries double the cost.

What Grok 4.6 means for AI agents

Competition is increasingly moving beyond models that simply generate answers towards systems that can complete longer workflows. For developers and businesses, the value of Grok 4.6 will ultimately depend on how reliably it can maintain context, use tools, identify mistakes and complete tasks without constant intervention.

If xAI's reported improvements translate into real-world performance, Grok 4.6 could strengthen its position in the growing market for AI agents, coding assistants and productivity tools.