Grok 4.7 is here, and xAI says it is its smartest model yet
Grok 4.7 has been introduced as xAI’s most capable model for coding, knowledge work and longer professional tasks, with improved safeguards and API access.
xAI has introduced Grok 4.7, calling it its most capable model yet for coding and knowledge-based tasks.
The model is aimed at developers and professionals handling tasks that can stretch across multiple steps, require careful checking and take considerably longer than a typical chatbot prompt.
Announced on September 21, 2026, Grok 4.7 is designed to work for longer on difficult problems, check its own output and handle larger amounts of context. It seems that xAI is trying to make Grok less of a quick-answer machine and more of a work partner.
Grok gets more time to think
Grok 4.7 uses a bigger base model than Grok 4.6 and has been trained for longer using reinforcement learning.
Reinforcement learning means the AI learns from feedback on how well it solves problems. For Grok 4.7, xAI says this training focused on more difficult problems that can take hours to solve instead of just a few minutes.
Grok 4.7 has also been trained to work directly with the tools and systems used by Grok Bot. That is intended to make it better at conversational tasks while helping it handle longer, more involved workflows.
Another focus is self-checking. Instead of simply producing an answer and moving on, Grok 4.7 is designed to spend more effort verifying its work, something that could matter when writing code, preparing documents or carrying out research.
The coding numbers are moving up
xAI's benchmark results point to improvements over Grok 4.6. On CursorBench 4.0, which tests longer software engineering tasks, Grok 4.7 scored 46.3%, compared with 40.4% for Grok 4.6. On DeepSWE v1.1, another software engineering benchmark, it recorded a high-effort score of 71.0%.
The improvements are not limited to coding. xAI says Grok 4.7 also performed better on tests covering legal work, healthcare reasoning, electrical engineering and multi-hour office tasks, including AA Briefcase, EEBench and the Harvey Legal Agent Benchmark.
That gives the model a broader target than programmers alone. Think research, analysis, presentations and other work where an AI needs to keep track of several moving pieces.
Safety is part of the pitch
xAI is also highlighting new safety measures. The company says Grok 4.7 is its strongest model yet at refusing unsafe requests and resisting jailbreak attempts. In cybersecurity tests, xAI says only 3.3% of risky dual-use prompts were allowed through on HackerBench v0.3, while legitimate security requests were rarely blocked.
Grok 4.7 is available through Cursor, the Grok API, Grok Build, model routers, third-party coding tools, and cloud platforms. Prices start at $2 per million input tokens and $6 per million output tokens. A faster version is also available at twice the speed for twice the price.
The bigger idea behind Grok 4.7 is simple. As AI moves deeper into professional work, being able to answer quickly is not always enough. Sometimes the useful assistant is the one that can keep working, check what it has done and come back with something that is actually ready to use.