SpaceXAI released the new Grok 4.6 model.

Grok 4.6 is designed for "long-lived" agents—systems capable of working on a given task for extended periods, autonomously restarting, and progressing toward their set goals. The model performs tasks in multiple stages, works with unfamiliar codebases, and tests its own results before advancing further.

These improvements resulted from a longer additional training cycle, the use of higher-quality engineering data, and expanded reinforcement learning for programming and intellectual labor.

In the Artificial Analysis Intelligence Index, the new version scored 61 points. It surpassed the Chinese open-weight model Kimi K3 from Moonshot and tied with GPT-5.6 Sol from OpenAI in maximum reasoning levels, trailing Anthropic's Claude Fable 5 Max by one point. Compared to Grok 4.5, released in July, results improved across all listed benchmarks, including CursorBench, FrontierCode, APEX-Agents, and Terminal-Bench.

Grok 4.6 outperforms its predecessor in code development, terminal operations, information handling, and agent performance. Meanwhile, the API pricing remains unchanged—$2 per million input tokens and $6 per million output tokens. This positions Grok 4.6 as a mid-range model compared to leading proprietary options and open-source solutions.

The model is already available in Cursor, Grok Build, and via the SpaceXAI, OpenRouter, Vercel, and Cloudflare APIs.

Source: SpaceXAI