SpaceXAI has released the Grok 4.7 model, which the company positions as its most advanced solution for programming and knowledge-based tasks. The new model features a larger base architecture and extended reinforcement learning cycles, focusing on long-context management, self-correction, and multi-step workflow execution.
Performance metrics show improvements across several benchmarks. In CursorBench 4.0, Grok 4.7 scored 46.3% compared to 40.4% for the previous version. Terminal-Bench 4.0 results rose from 20.3% to 38.0%. Professional domain benchmarks also saw growth: EEBench reached 64.0%, Harvey Legal Agent Benchmark hit 19.6%, and HealthBench Professional climbed to 56.7%.
The model includes a new security framework, achieving a 62.4% score on the LatchBio biosafety benchmark. In HackerBench v0.3, it restricted high-risk dual-use prompts to 3.3%. SpaceXAI has also initiated an invite-only red teaming program for cybersecurity partners to focus on defensive research.
Grok 4.7 is now available via Cursor, Grok Build, the Grok API, and various cloud platforms. Despite the performance upgrades, the pricing remains unchanged from Grok 4.6, starting at $2 per million input tokens and $6 per million output tokens.
