September 22nd news, yesterdaySpaceXAI Official Release Grok 4.7. Main programming, long missions and professional expertise. The new model has now entered Cursor, Grok Build and Grok API and will be provided through a third-party programming tool, model route platform and cloud service。
According to the official presentation, Grok 4.7 used a new, larger base model than Grok 4.6 and extended the length of intensive learning and training, with a focus on strengthening context management, self-censorship and the ability to perform complex tasks for several hours。
In the baseline test, Grok 4.7 xHigh obtained 46.3% on CursorBench 4.0, above 40.4% in Grok 4.6 High; DeepSWE v1.1 from 652% to 71.0%, Terminal-Bench 4.0 from 20.3% to 38.0%。
However, in a horizontal comparison, the new model is not fully ahead: CursorBench 4.0 is still below Fable 5.1 Max and DeepSWE v1.1 is slightly below GPT-5.6 Sol Max。
Prices are a major competitive point in this escalation. Grok 4.7 API pricing consistent with previous generation, per million input token 2 united states dollars, token 6 dollars; and a fast version with about double output speed, with the same double price。