Grok 4.7 released
On 21 September 2026 SpaceXAI released Grok 4.7, which it calls its most capable model for coding and knowledge work, in Cursor, Grok Build and its API as grok-4.7. It uses a new, larger base model than Grok 4.6 and a longer reinforcement-learning run on tasks that take many hours; price and speed are as for Grok 4.6: $2 and $6 per million input and output tokens up to 200K prompt tokens.
Why it matters
Six weeks after 4.6 the company changed not only the post-training but the base model. Its headline claims twice the speed at half the price of comparable models; in the same tables Grok 4.7 trails Fable 5.1 on several tests.
API and prices. The API release notes ('September 21'): grok-4.7, a 500K-token context, text and image input, text output with no length limit; $2 / $0.50 / $6 per million input, cached input and output tokens up to 200K tokens and $4 / $1 / $12 above; reasoning efforts low, medium, high (default), xhigh. Grok 4.7 Fast costs twice as much and is available only in Cursor and Grok Build, not in the public API. The weights are not open. The company's figures. CursorBench 4.0: 46.3% against 40.4% for Grok 4.6, 41.7% for GPT-5.6 Sol and 51.8% for Fable 5.1. DeepSWE v1.1: 71.0% (at high effort), GPT-5.6 Sol 72.7%, Fable 5.1 70.0%. Terminal-Bench 4.0: 37.6% against 37.3% for GPT-5.6 Sol; the Fable 5.1 column shows 57.9%, while Anthropic's pages give Fable 5.1 55.8% and give 57.9% for GPT-6 Astra, so the label in the company's table may be a slip; the record does not resolve it. GDPval: 1695 against 1735 for Fable 5.1 (max). Safety: HackerBench v0.3 lets 3.3% of risky prompts through; LatchBio biosafety 62.4%; select cybersecurity partners get invite-only access to red-team capabilities. What the record does not claim. All figures are the company's own; no independent measurements were read. The post is dated 21 September 2026, the same date as the API release notes.