xAI has launched Grok 4.7, its latest large language model designed for complex coding and knowledge-intensive tasks. The model extends processing duration for challenging problems, incorporates more rigorous self-verification steps, and deploys a refined safety framework described as the most robust to date. Pricing and deployment speed match those of its predecessor, Grok 4.6, maintaining accessibility for existing users.

Performance evaluations on CursorBench 4.0, a benchmark emphasizing extended coding workflows, position Grok 4.7 at the leading edge of cost-efficiency within its category. The model demonstrates marked improvements in verifying its own outputs and sustaining contextual awareness across longer interactions. Training involved an expanded base architecture and an extended reinforcement learning phase focused on multi-hour problem solving, alongside specialized instruction in interpreting the Grok Bot interface to enhance conversational fluency and factual accuracy.

Benchmark results indicate stronger performance across professional simulation tasks. In GDPval and AA Briefcase assessments—where AI handles document creation and presentation development for roles such as legal and financial analysts—Grok 4.7 surpasses Grok 4.6 and matches the output quality of competing frontier models. Additional testing on DeepSWE v1.1, EEBench, and Terminal-Bench 4.0 further confirms gains in software engineering and technical task execution, with notable reductions in error rates compared to prior versions.

Safety protocols represent a central advancement. Grok 4.7 exhibits heightened resistance to misuse while preserving utility in legitimate applications across cybersecurity and biomedical domains. It achieved the top score on LatchBio’s biosafety evaluation at 62.4% success in benign task execution with appropriate refusal of hazardous prompts. On HackerBench v0.3, designed to assess risky dual-use scenarios, the model permitted only 3.3% of potentially dangerous inputs to proceed, while maintaining minimal interference with authorized security research activities. Selected cybersecurity researchers now access limited red-team capabilities through invitation-only channels.

The model enters general availability immediately through Cursor and Grok Build platforms, with API access also offered via third-party integration points and cloud service providers. Input pricing begins at $2 per million tokens, with output priced at $6 per million tokens. A high-speed variant provides double the output throughput at double the cost. Early access is available through the Grok Build portal for prospective users.