An overview of Grok 4.7 and its benchmark performance

Grok 4.7 is a new artificial intelligence model built with a larger base model than its predecessor, Grok 4.6. The training process for this version included a longer reinforcement learning run, which is a method of teaching models through trial and error using rewards and penalties. This training focused on a harder mix of tasks that take many hours to complete.

Model improvements and performance

The model features improved self-verification, meaning it is better at checking its own work and managing longer context windows. It has also been trained to natively understand the Grok Bot harness, which helps with conversational tasks and general knowledge work. On CursorBench 4.0, a benchmark that stresses longer-running coding tasks, Grok 4.7 delivers strong price-performance.

The model is also better at creating documents and presentations. In professional benchmarks like GDPval and AA Briefcase, where artificial intelligence is asked to work on tasks handled by lawyers, nurses, and financial analysts, Grok 4.7 improves upon Grok 4.6 and performs comparably to other frontier models.

Safety and cybersecurity

Grok 4.7 was built with an entirely new safeguard stack, which is a collection of safety filters and checks. It is designed to be strong on refusals and jailbreak resistance. In dual-use domains like cybersecurity and biological work, it leads on utility for benign tasks while maintaining safe refusals on dangerous ones.

The model tops LatchBio’s biosafety benchmark at 62.4 percent and shows high safety on HackerBench v0.3, allowing only 3.3 percent of risky dual-use prompts through while rarely blocking legitimate security work. Select cybersecurity partners now have invite-only access to red-team capabilities for defense research.