Grok 4.6 Released

xAI released Grok 4.6, focusing on long-running agents and complex interactive/visual tasks. It excels at continuously performing multi-step research, information analysis, codebase work, and creating applications and business deliverables.

Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index, tying with GPT-5.6 Sol. It recorded 1753 on GDPVal-AA v2, 69.9% on CursorBench v3.2, and 65.9% on DeepSWE v1.1; competitor model figures are based on system cards and benchmark leaderboards published by each developer.

The model underwent additional training longer than Grok 4.5. After strengthening its foundation with reasoning, advanced technical concepts, and high-quality engineering data, it reconstructed SFT trajectories generated by Grok 4.5 across various domains such as STEM, software engineering, and knowledge work, and filtered them through model-based inspection.

Reinforcement learning included domain-specific agent environments such as kernel optimization, web development, and CAD, in addition to general coding and knowledge work. It can structure product ideas, implement core interactions, and improve by reflecting feedback multiple times; during long task processes, it also performs self-testing and verification.

Grok 4.6 is currently available on Cursor and Grok Build. During the first week of release, included usage is doubled on both services.