SpaceXAI has officially released Grok 4.7, positioning it as its strongest model for coding and knowledge work, now available in Cursor, Grok Build, and the Grok API. The company says it shows clear gains over Grok 4.6 on benchmarks like CursorBench 4.0, DeepSWE v1.1, and Terminal Bench 4.0, with API pricing unchanged and a focus on long-horizon execution, self-checking, and professional tasks.
SynthePulse Insight · AI deep readingMembers
Version 1 · 5 sources
SpaceXAI released Grok 4.7 on September 21, 2026, claiming it is stronger at coding and knowledge work, with pricing and speed on par with Grok 4.6. But Chinese-language coverage and developer feedback point to two contrasts at once: it does not lead across the board in cross-model comparisons, and completing real tasks may consume more tokens.
SpaceXAI officially released Grok 4.7 on September 21, 2026, positioning it as its strongest model for coding and knowledge work and claiming it can be twice as fast as comparable models at half the price. The model has entered Cursor, Grok Build and the Grok API, and will also be offered through third-party coding tools, model routing platforms and cloud services. Standard API pricing is consistent with Grok 4.6: $2 per million input tokens and $6 per million output tokens; there is also a fast version with double the output speed and a correspondingly doubled price.
The focus of this upgrade is not the chat experience or multimodal display, but long-duration execution, self-checking and professional knowledge work. SpaceXAI says Grok 4.7 uses a larger base model than Grok 4.6, with longer reinforcement learning and harder training tasks, a considerable portion of which take hours to complete; the model also strengthens long-context management and result verification, and was natively trained for the Grok Bot runtime environment.
Musk publicly stated that Grok 4.7 achieves a very competitive balance among intelligence level, operating speed and cost. This statement comes from his tweet and is the publisher's position, not an independently verified conclusion.
Multiple officially published results show that Grok 4.7 has made clear progress compared with Grok 4.6: CursorBench 4.0 rose from 40.4% to 46.3%, Terminal-Bench 4.0 from 20.3% to 38.0%, EEBench from 53.0% to 64.0%, Harvey Legal Agent Benchmark from 15.8% to 19.6%, and GDPval Elo from 1605 to 1695. Machine Heart also mentions that AA Briefcase v1.1 rose from 1546 points to 1657 points and HealthBench Professional from 48.5% to 56.7%, while APPSO mentions that DeepSWE v1.1 improved from 65.2% to 71.0%. These numbers all come from the internal comparison table published by SpaceXAI; they can show the change between the old and new generations, but cross-model comparisons are still affected by reasoning intensity, tool configuration and the testing environment.
After horizontal comparison, Grok 4.7's advantages are mainly concentrated in price and some professional tasks, and its overall results do not lead across the board. APPSO points out that on CursorBench 4.0 and Terminal-Bench 4.0 it is below Fable 5.1 Max, and on DeepSWE v1.1 it is slightly below GPT 5.6 Sol Max; however, in Harvey Legal Agent Benchmark it scored 19.6%, clearly higher than several comparison models in the official table. Machine Heart also mentions that Grok 4.7 is close to Fable 5.1's 1735 points on GDPval, and higher than GPT-6 Astra's 1542 points.
Artificial Analysis subsequently gave it an Intelligence Index score of 46, saying that Grok 4.7 puts SpaceXAI into the world's top four AI labs, and that its Coding Agent Index performance also exceeds GPT 5.6 Sol. This third-party evaluation result comes from APPSO's account and can serve as an external reference, but it should not be equated with a comprehensive capability ranking.
This article is mainly based on SpaceXAI's official release information, reports by Machine Heart and APPSO, commentary by TechRepublic, and developer and user feedback on the X platform. Official benchmarks and pricing are the publisher's account; cross-model comparisons are affected by reasoning intensity, tool configuration and the testing environment; developer tests have limited samples; claims about the reasons for delays, parameter scale and training data come from secondhand accounts and have not been confirmed by first-party sources.
Primary report
Primary source
Also covered by 4 sources