Grok 4.7

Grok 4.7, released September 21, 2026, was the model Elon Musk had trailed for most of the month: a larger successor to Grok 4.6 that he said traded some serving speed for capability, and that he claimed had been trained in part on SpaceX's own engineering data. It shipped with a 500,000-token context window and at Grok 4.6's prices, unchanged in every column — $2 per million input tokens and $6 per million output under 200K, doubling to $4 and $12 once a prompt crosses that threshold, with cached input at $0.50 and $1.00. The launch came in two pieces that did not compare the same models: a post on X setting Grok 4.7 at xHigh against Grok 4.6, GPT-5.6 Sol and Claude Fable 5.1, and a launch page whose charts swapped in GPT-6 Astra, the OpenAI flagship that had shipped three weeks earlier.

The engineering claim was the one xAI built the launch around, and the two surfaces answered it differently. Grok 4.7 scored 64.0% on EEBench, an electrical-engineering set, against 53.0% for Grok 4.6 — eleven points, its largest gain over its predecessor anywhere — and ahead of Claude Fable 5.1's 56.4%, but the launch page's own chart put GPT-6 Astra above it at 69.3%. The knowledge-work rows fell the same way, second to a rival in each case: a GDPval-AA v2.1 Elo of 1695 against 1605 for Grok 4.6 and 1735 for Claude Fable 5.1, and 1657 on AA-Briefcase v1.1 against 1546 and 1678. Where xAI led outright was legal work, at 19.6% on Harvey's Legal Agent Benchmark — up from 15.8% for Grok 4.6 and far above the 2.5% and 6.7% of the models beside it on the X table. Coding and terminal work stayed the softer side: 71.0% on DeepSWE v1.1, footnoted as run at High rather than xHigh effort, sat just behind GPT-5.6 Sol's 72.7%, and 38.0% on Terminal-Bench 4.0 trailed Claude Fable 5.1's 57.9% by twenty points, with 56.7% on HealthBench Professional third of the four columns. As at the Grok 4.6 launch, the figures for rival models were xAI's own presentation rather than each lab's published results.