SpaceXAI has launched Grok 4.7, enhancing its programming capabilities and achieving improved performance metrics, while maintaining pricing.
On September 21, SpaceXAI introduced Grok 4.7, a newly upgraded model aimed at enhancing programming, agent tasks, and knowledge management capabilities.
The latest iteration boasts a more robust foundational model than its predecessor, Grok 4.6. Improvements include extended reinforcement learning periods and the ability to tackle more complex tasks, some requiring several hours to complete.
Performance Metrics
In recent internal tests using CursorBench 4.0, Grok 4.7 achieved a score of 46.3%, marking an increase from Grok 4.6's previous score of 40.4%. For comparison, other models like GPT-5.6 Sol scored 41.7%, while Claude Fable 5.1 led with 51.8%.
In the AA Briefcase test, focusing on longer tasks, Grok 4.7 attained an Elo score of 1657, compared to 1546 for Grok 4.6. Additionally, the model's performance in EEBench improved from 53% to 64%, and in the Harvey Legal Agent Benchmark, it rose from 15.8% to 19.6%.
According to an independent analysis by Artificial Analysis, Grok 4.7 earned 46 points on the Intelligence Index, slightly ahead of Grok 4.6. When paired with Grok Build, its performance on the Coding Agent Index reached 56 points, up from 47.
Despite these advancements, Grok 4.7 requires significantly more tokens for complex tasks, averaging around 81,000 output tokens in xhigh mode, compared to 36,000 for the previous version.
Pricing and Features
Pricing for Grok 4.7 remains consistent, with charges of $2 per million input tokens for requests up to 200,000 tokens, $0.50 for cached inputs, and $6 for output tokens. For larger requests, the costs double.
The model can handle up to 500,000 tokens in context and is capable of processing both text and images. It offers four reasoning depth settings: low, medium, high, and xhigh. A new variant, Grok 4.7 Fast, features twice the generation speed at double the cost, available exclusively through Cursor and Grok Build, not via the public API.
Moreover, SpaceXAI has revamped its protection system, demonstrating a decrease in dangerous query passes from 3.3% in HackerBench v0.3 to a score of 62.4% in the LatchBio biosecurity test. The previous model, Grok 4.6, was released just a month earlier, maintaining similar pricing structures and capabilities.
