The September 21, 2026 release builds on a larger base than Grok 4.6, with extended training on harder multi-hour tasks and better self-verification — at the same 500K context and the same headline price.
2026-09-21 · by Moe Ameen
xAI released Grok 4.7 on September 21, 2026, calling it its most powerful model for coding and knowledge work — one it frames as roughly twice as fast at half the price of comparable models. It builds on a larger base model than Grok 4.6 (Elon Musk publicly put the new base at around 2.1 trillion parameters, up from 4.6's ~1.5 trillion), trained with extended reinforcement learning on harder, multi-hour tasks. xAI says the model works longer on difficult problems, checks its own work more carefully, and manages long context better than 4.6, and that it was trained to natively understand the Grok Bot harness for conversational and general knowledge-work tasks.
On xAI's launch benchmarks, 4.7 improves over 4.6 across coding, agentic, and knowledge evals: CursorBench 4.0 rises to 46.3% (from 40.4%), DeepSWE v1.1 to 71.0% at high effort (from 65.2%), Terminal-Bench 4.0 to 38.0% (from 20.3%), EEBench to 64.0% (from 53.0%), HealthBench Professional to 56.7% (from 48.5%), and the Harvey Legal Agent benchmark to 19.6% (from 15.8%). These are xAI's own figures; independent evaluations had not landed at launch.
Grok 4.7 keeps the 500,000-token context window and the headline API pricing from 4.6 — $2 per million input tokens, $0.50 per million cached input, and $6 per million output — with a faster variant at roughly double the speed and double the price. As with 4.6, that headline rate doubles once a single request crosses 200,000 tokens ($4 input, $12 output). It ships with what xAI describes as its best-calibrated safeguards to date, and is available through the xAI API, Grok Build, Cursor, and third-party coding harnesses and cloud platforms.
The move on a release like this is to put the new model where it's strongest and hand off the instant it stops. Grok 4.7's gains are in reasoning over hard, long problems and checking its own work — so it's a better place than ever to digest a dense source (a quarterly report, a long transcript, a research thread) and draft from it. What it still won't do is make or ship any of it: no avatar video, no carousel, no captions, no schedule, no publishing.
That's the handoff Kompozy is built for, and you can act on it the same day the model ships. Take Grok 4.7's draft or research summary and drop it into Kompozy as a source. Kompozy generates the finished formats it implies — a HeyGen persona short with burned-in captions, a brand-exact HyperFrames carousel, quote cards, a blog draft, a newsletter — all held to your voice through a Persona Brief, then schedules and autopilots the set across eight social platforms plus blog and email from one queue. You don't wire Grok into Kompozy; it runs its own generation on managed Claude and OpenAI models, so 4.7's output goes in as raw material and the engine ships it. Even covering this launch is a use case: draft the explainer in Grok, then produce and publish the whole content set in Kompozy.
Grok 4.7 is xAI's flagship model, released September 21, 2026, positioned as its most powerful model for coding and knowledge work. It reads text and images and returns text, keeps a 500,000-token context window, and is built on a larger base than Grok 4.6 with extended training on harder, multi-hour tasks and better self-verification.
Via the xAI API it is $2 per million input tokens, $0.50 per million cached input, and $6 per million output — the same headline rate as Grok 4.6. That rate doubles once a request crosses 200,000 tokens ($4 input, $12 output), and a faster variant runs at roughly double the speed for double the price.
A larger base model, extended reinforcement learning on harder multi-hour tasks, better self-checking and long-context handling, and native understanding of the Grok Bot harness. xAI reports gains across its coding, agentic, and knowledge-work benchmarks at the same 500K context and same price.
It drafts the text — scripts, captions, angle batches, research summaries — quickly and well, but it generates no images, video, or audio and publishes nothing. To turn its drafts into finished, scheduled posts you pair it with a content engine like Kompozy that produces the media and publishes across platforms.