Performance and Capabilities
xAI officially launched Grok 4.7 on September 21, 2026, targeting developers and enterprises focused on coding, knowledge management, and long-running agent workflows, NeoTeo reported. The model introduces a larger base architecture and extended reinforcement learning, which xAI claims improves the model’s ability to manage multi-step assignments and self-verification tasks.
In official benchmarks, Grok 4.7 shows notable gains over its predecessor, Grok 4.6. On the Terminal-Bench 4.0 evaluation, the model reached 38.0%, nearly doubling the 20.3% score of the previous version. It also saw improvements on the CursorBench 4.0 and the Harvey Legal Agent Benchmark. However, the model does not lead across all categories; TechRepublic notes that competitors like Fable 5.1 Max outperform Grok 4.7 on several coding and clinical reasoning tasks, suggesting that the model’s utility is highly workload-dependent.
Cost Considerations and Token Usage
While the API pricing remains fixed at $2 per million input tokens and $6 per million output tokens—consistent with Grok 4.6—the total cost of ownership remains a point of scrutiny for developers. Independent analysis by Artificial Analysis indicates that Grok 4.7 at its ‘xHigh’ reasoning setting consumes approximately 81,000 output tokens per task, significantly higher than previous iterations and some rival models like GPT-6 Astra.
Because of this high token consumption, organizations are advised to evaluate the model based on the cost per successful job rather than raw API rates. xAI has also introduced a ‘fast’ variant for users requiring higher output speeds, though this option carries double the standard pricing.
Safety and Accessibility
Grok 4.7 incorporates a new safety stack designed to improve jailbreak resistance while reducing unnecessary refusals. xAI reports a 62.4% score on LatchBio’s biosafety benchmark. Currently, the model is available through various channels, including Cursor, the Grok API, and third-party coding harnesses. The company is also granting invite-only access to select cybersecurity partners for red-team testing to further refine these safeguards.