SpaceXAI has released Grok 4.7, which it describes as its strongest model so far for writing software and knowledge work. The company says the model can stay with difficult problems for longer, check its own answers more carefully, and follow safety rules more reliably than Grok 4.6. It is sold at the same starting price and the same standard speed as Grok 4.6. SpaceXAI also says it is about twice as fast and about half as expensive as some rival models in the same class.
Grok 4.7 uses a new, larger base model than Grok 4.6. It received a longer period of reinforcement learning. That extra training focused on work that can take many hours, such as large coding jobs and multi-step office tasks. The model accepts text and images and replies with text. It can keep about 500,000 tokens of material in mind at once. Users can choose how hard the model thinks. Official documents list a knowledge cutoff of May 2026, meaning later events are not part of its training data.
On company tests, Grok 4.7 scored higher than Grok 4.6 on coding, electrical engineering, legal work, clinical reasoning, and long office tasks. Results against other leading models were mixed. It led some exams, including an electrical-engineering test and a legal-agent test, while other models still led on some long coding and health tests.
Safeguards, price, and where it can be used
SpaceXAI says Grok 4.7 uses a rebuilt safeguard system. The company reports strong scores on tests that measure whether a model refuses dangerous biology and cybersecurity requests while still helping with lawful security work.
Standard use costs two dollars per million input tokens and six dollars per million output tokens. A faster version, sold only in some coding products, costs twice as much and produces replies about twice as quickly. Grok 4.7 is available in the coding tools Cursor and Grok Build, through the Grok API.