SpaceXAI introduces Grok 4.7 language model

The novelty is built on a larger model compared to Grok 4.6. During its development, engineers increased the duration of reinforcement learning and placed an emphasis on tasks requiring hours of continuous computing. The model received a 500,000-token context window, learned to verify output results in greater detail, and gained built-in support for the Grok Bot environment.

According to the company, the model scored 46.3% on the CursorBench 4.0 coding test and achieved a result of 71% in the DeepSWE v1.1 benchmark. In the GDPval professional tasks test, Grok 4.7 scored 1,695 points. The safety mechanisms system was also overhauled: in the HackerBench v0.3 cyberthreat test, the proportion of missed hazardous prompts was 3.3%, while in the LatchBio biosecurity trial, the figure reached 62.4%.

The base API pricing remained at the level of the previous version at $2 per 1 million input tokens and $6 per 1 million output tokens. In addition, a Fast version with double the generation speed was introduced at twice the rate. The model is available via the API, as well as in Cursor and the Grok Build environment.