SpaceXAI has released a new version of its cutting-edge Grok 4.6 model, which it says is on par with OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5 AI models.
Image source: SpaceXAI
According to the company, Grok 4.6 is designed for “long-lived” agents that can complete assigned tasks for long periods of time, restart independently, and move toward predetermined goals. According to reports, Grok 4.6 can participate in multiple stages of tasks, use unfamiliar code libraries, and test its own results before continuing.
In the artificial intelligence analysis intelligence index test, the new version of the AI model scored 61 points, surpassing China’s popular Moonshot open weight model – Kimi K3 – with the highest inference level equivalent to the competitor GPT-5.6 Sol, one point behind Fable 5 Max. Compared to Grok 4.5, released in July, the new version is smarter in all listed tests, including CursorBench, FrontierCode, APEX-Agents and Terminal-Bench.
Increased capabilities are the result of longer incremental training cycles, better engineering data, and enhanced programming and knowledge reinforcement learning.
More importantly for the company, Grok 4.6 shows significant improvements over its predecessor in code development, terminal operation, information processing and agent performance, while maintaining API prices of $2 per 1 million input tokens and $6 per 1 million output tokens. Sources note that this makes it a mid-priced model compared to leading options around the world, both proprietary and open source. venture capital.
Grok 4.6 is now available in Cursor, Grok Build, and through the SpaceXAI, OpenRouter, Vercel, and Cloudflare APIs.
If you find an error, select it with your mouse and press CTRL+ENTER.
