Grok 4.7 is now available on VEGA

Grok 4.7, the new flagship model from SpaceXAI, is now live on VEGA. You can pick it in the model selector under the Pro category for 3 stars per request, the same price Grok 4.6 had. It is also part of the ProCoder tier in Code mode.

The model is built for work that takes hours: long agent runs, multi-step coding, large documents. SpaceXAI moved it to a larger base model, about 2.1 trillion parameters versus 1.5 trillion in Grok 4.6, and extended reinforcement learning on long tasks. The company says the model now checks its own work more reliably and loses the thread less often. The context window stays at 500K tokens, input can include text and images, and reasoning effort is adjustable across four levels from low to xhigh.

Benchmarks paint a mixed picture. In SpaceXAI's own tests Grok 4.7 beats GPT-5.6 Sol in five of seven evaluations, with the clearest gains in engineering and legal tasks, though Grok ran at xhigh effort against competitors at lower settings. Independent measurements from Artificial Analysis place it mid-table, and terminal-based coding remains its weak spot.

Part of the training data comes from SpaceX itself: Starlink satellite telemetry, production records, and engineering failure logs. The idea is a model that reasons about hardware and physical systems better than one trained on web text alone. Grok 4.7 also ships with a rebuilt safety stack and, per the company's testing, holds up well against jailbreak attempts.

API pricing stays at the Grok 4.6 level. A faster variant, Grok 4.7 Fast, exists only inside Cursor and Grok Build for now, so on VEGA you get the standard version. Open the model picker and choose Grok 4.7 to try it.