xAI Drops Grok 4.7: 2.1 Trillion Parameters, Twice the Speed, Rebuilt Safety

xAI shipped Grok 4.7 on September 21, 2026, its most capable model to date. The headline number is 2.1 trillion parameters—but the more interesting story is what xAI did with the training run to make those parameters useful for the kinds of tasks that actually matter at work.

What's New in Grok 4.7

According to xAI's release notes, Grok 4.7 uses a new, larger base model with a longer reinforcement learning run and training that deliberately weights difficult multi-hour tasks. The model features a 500K context window, text and image inputs, and a reasoning effort dial (low, medium, high, or xhigh) that lets users tune latency versus depth.

xAI claims it runs twice as fast as comparable models at this scale. For agentic workflows where latency compounds across dozens of tool calls, a genuine 2× speed improvement translates directly to wall-clock time saved on complex tasks.

Benchmark Improvements

The published numbers show consistent gains over Grok 4.6 across xAI's full benchmark suite. CursorBench 4.0 improved from 40.4% to 46.3%. DeepSWE v1.1 jumped from 65.2% to 71.0%. EEBench climbed from 53.0% to 64.0%. Terminal-Bench 4.0 nearly doubled, from 20.3% to 38.0%—the largest relative gain and the one that matters most for agentic coding workflows.

SpaceX Engineering Data and Rebuilt Safety

Two less-discussed details: Grok 4.7 was trained on SpaceX engineering datasets, which explains some of the improvements in technical and scientific domains. xAI also completely rebuilt its safeguard system, which it says is now more resistant to jailbreaks while generating fewer unnecessary refusals. That combination is harder to achieve than it sounds, and most safety updates don't actually land both sides of the tradeoff.

Pricing and Access

API pricing starts at $2 per million input tokens and $6 per million output tokens—unchanged from Grok 4.6, despite the significantly larger model. Access is available through Cursor, Grok Build, and the xAI API directly. Function calling, web search, X search, and code execution are all supported.

For teams already using Grok 4.6 in production, upgrading to 4.7 at the same price with meaningfully better agentic benchmark performance is a straightforward decision. For teams evaluating the xAI ecosystem for the first time, the Terminal-Bench jump is the number to watch—it is the closest available proxy to real-world, long-horizon task performance.