xAI Releases Grok 4.7: An Improvement Over the Previous Version at the Same Low Price

xAI (branded on the page as SpaceXAI) released Grok 4.7, the newest version of its Grok model. Versus the previous 4.6, it promises a larger base model, better at verifying its own work, better long-context handling, and an entirely new safeguard layer.

In brief

• What shipped: Grok 4.7, on September 21. Versus 4.6: a larger model, a longer reinforcement-learning run on harder tasks, native understanding of the Grok Bot harness, and a rebuilt safeguard stack.
• Price and access: $2 input / $6 output per million tokens (same as 4.6; the fast variant is twice the speed at twice the price). It claims to be "twice as fast, at half the price of comparable models." Available via Cursor, Grok Build, the Grok API, third-party coding harnesses, model routers, and cloud platforms.
• Comparison (xAI's own measurements): it beats 4.6 across every shared benchmark. Against frontier models (GPT-5.6 Sol, Fable 5.1) the picture is mixed — ahead on some (EEBench, Harvey Legal, and above GPT on AA Briefcase), behind on others (Fable 5.1 leads clearly on Terminal-Bench 4.0 and CursorBench; it trails both on HealthBench). So it's competitive, but not "first at everything."
• Safety note: the page also shares safety measurements — it reports that a "dual-use" test let through about 3.3% of risky prompts, and it scored 62.4% on a biosafety benchmark. These, too, are xAI's own measurements.

Our take

The story runs at the same rhythm as Grok 4.6 in #33: a clear step over its predecessor, and a mixed picture against frontier models — ahead on some tests, behind on others (notably where Fable 5.1 leads clearly on Terminal-Bench and CursorBench), all on xAI's own numbers. For our audience the real value is again price and speed: $2/$6 and a "twice as fast, half the price" claim make it cheap for a frontier-competitive model, and it's widely available. If you do coding or agent work, it's a low-cost option worth trying. The honest caveats are the same: the numbers aren't independent (real comparisons come within weeks), and "competitive" isn't "best." One more note: this is the third Grok version we've covered in about 2.5 months (4.5, 4.6, 4.7) — the real signal may not be any single model, but how often frontier labs now ship these cheap, fast incremental releases.

Kaynak: xAI