xAI Ships Grok 4.7 to Enable Longer Agent Runs

xAI has launched Grok 4.7, a reinforcement-learning upgrade to Grok 4.6, arriving just two months after the previous version. The release aims to improve performance on complex tasks by allowing the model to reason for longer periods, verify its own outputs, and utilize enhanced safety safeguards. This update addresses previous reinforcement-learning bugs where internal builds finished difficult assignments too quickly without checking their work. To fix these issues, Grok 4.7 allocates additional internal computation tokens to process extended reasoning before delivering a final answer.

The model is accessible immediately through the xAI API, Cursor, and Grok Build. Standard pricing holds steady at $2 per million input tokens and $6 per million output tokens, with another tier noted at $4 per million input and $12 per million output tokens. Although xAI has not revealed the model's exact parameter count or base design—leaving reports of a 2.1-trillion-parameter model unconfirmed—its evaluations target agentic benchmarks such as DeepSWE, CursorBench, Terminal-Bench, APEX, and AA-Briefcase. In an internal demonstration, Grok 4.7 built a more complete open-world city game than Grok 4.6 within the same timeframe, though independent evaluations are still needed.

For developers, the transition to Grok 4.7 requires few API changes, but the model's extended reasoning capabilities alter production economics. Because the system spends more time thinking, it can generate higher output volumes and increase latency, raising the total cost per task even though token rates remain identical. Teams should run canary deployments to evaluate performance on multi-step agentic workloads, such as repository-level coding, multi-file user-interface generation, and sequential tool use. Measuring the cost per successful task, rather than simple token rates, will help teams determine if the upgrade delivers genuine operational value.