The new Grok 4.7 coding model from xAI, released on September 21, 2026, has secured the top three positions on the VulcanBench Frontier v4 leaderboard. The model allows users to adjust reasoning effort, with the extra-high, high, and medium settings achieving scores of 93.15, 92.71, and 92.30 respectively.
These results place Grok 4.7 ahead of its closest competitor, Claude Fable 5.1, which recorded a score of 91.84. At the extra-high effort level, the model successfully completed all 23 behavioral-reconstruction tasks, which require software to be rebuilt to match the behavior of a reference model.
Grok 4.7 features an expanded base model, supports text and image inputs, and offers a context window of 500,000 tokens. The company has maintained the pricing structure of its predecessor, Grok 4.6, at $2 per million input tokens and $6 per million output tokens.
Beyond VulcanBench, xAI reports performance improvements over Grok 4.6 on benchmarks such as CursorBench and Terminal-Bench. The developers suggest that the tiered reasoning effort allows users to optimize costs by selecting lower settings for routine tasks while reserving higher settings for complex engineering problems.
