I’ll verify public reporting on a Grok 4.7 release today so that the article remains up to date. xAI released Grok 4.7 on 21 September 2026 as its most powerful model to date for coding and knowledge work; the announcement said, a larger successor to Grok 4.6 that stays on harder jobs longer, double-checks its work more thoroughly, ships with tighter safety rails, and maintains the price and serving speed of its predecessor.
The firm claimed Grok 4.7 is trained on a new, bigger core model and a longer RL run on more difficult tasks, including long-horizon tasks that take hours to complete. ‘This training improved long-term self-verification, long-horizon context handling, and capabilities inside agent environments like Grok Build and Grok Bot, ‘ xAI said. Multiple reports also mentioned 500,000 token context window, text and image inputs with text output, knowledge cut-off in ‘H1 2026’.
On xAI’s exam scores, the model was up overall against Grok 4.6. Cursor Bench 4.0, which is more focused on heavy software-engineering content, increased from 40.4 percent to 46.3 percent. Terminal-Bench 4.0 increased from 20.3 percent to 38.0 percent on the company’s harness, DeepSWE v1.1 reached 71.0 percent at high effort, and EEBench was reported at around 64 percent. Most independent coverage regarded those results as substantial but not enough to beat present top performers. Claude Fable 5.1 and GPT-6 Astra still sat atop at some headline assessments, including certain sections of GDPval, AA-Briefcase, and a few agentic coding evaluations.
You Might Be Interested In:
- Apple Launches iOS 27 with AI-Powered Siri
- How to Accept Stablecoin Payments on Stripe: A Step-by-Step Guide
- OpenAI Launches GPT-6 Astra, Calling It a Generational Leap Toward AGI
- Building Business Resilience in an Uncertain World
Pricing is the second part of the pitch. Basic API charges begin at $2 and $6 per million input and output tokens respectively, which is the same as Grok 4.6, with cached input listed at $0.50, and pricing doubling following the introduction of a prompt longer than 200,000 tokens. A fast version, with double output rate for double price, was available at launch in Cursor and Grok Build rather than via the public API. When offering, xAI claimed it was twice as fast and half as expensive per unit of price-performance measured against respective frontier models.
Availability was immediate in developer channels. Grok 4.7 went live in Cursor, Grok Build, and the Grok API, while third-party coding harnesses, model routers, and cloud platforms began to gain access. Safety was another focus: xAI said the model is leading on some dual-use assessments, like Latch Bio’s biosafety benchmark at 62.4 percent, by remaining effective on safe work and refusing safe requests more consistently. The launch came after a run of several public delays, and the early consensus is that Grok 4.7 is a substantive upgrade at an aggressive price point, even if it’s still chasing the very top of the frontier.
