The August 12, 2026 release is an incremental upgrade over Grok 4.5 — tuned for long-running agents, coding, and interactive prototypes — at the same 500K context and the same headline price.
2026-08-12 · by Moe Ameen
xAI released Grok 4.6 on August 12, 2026, positioning it as an incremental upgrade over Grok 4.5 rather than a new generation. The release is aimed at long-running agents: multi-step tasks the model can hold and self-verify, plus coding and turning a product idea into a working interactive prototype. xAI says 4.6 produces stronger first passes on visual and interactive projects than 4.5 did, crediting longer supplemental training on curated model-generated data, SFT trajectories regenerated with Grok 4.5, and reinforcement learning on agentic tasks across coding, knowledge work, and domain-specific environments.
On xAI's launch benchmarks, Grok 4.6 improves across the board. The Artificial Analysis Intelligence Index rises to 61, up from 56 for Grok 4.5 and level with GPT-5.6 Sol on that composite. Coding and agent evals moved more sharply — DeepSWE v1.1 to 65.9% (from 54.0%) and APEX-Agents to 57.5% (from 47.1%) — alongside GDPVal-AA v2 at 1753, CursorBench v3.2 at 69.9%, and FrontierCode v1.1 at 61.3%. These are xAI's own figures; independent evaluations had not landed at launch.
The model keeps Grok 4.5's 500,000-token context window and adds an "xhigh" reasoning level for the hardest tasks. API pricing is unchanged at the headline — $2 per million input tokens, $0.50 cached, and $6 per million output — but a long-context band doubles the rate to $4/$12 once a request crosses 200,000 tokens, and a faster variant costs about double the standard rate. Grok 4.6 is available through the xAI API (model string grok-4.6), Grok Build, Cursor on all plans, OpenRouter, Vercel, and Cloudflare, with 2x included usage in Grok Build and Cursor for the first week.
The practical move on a release like this is to use the new model where it's genuinely stronger and hand off the moment it stops. Grok 4.6's upgrade is agentic stamina — it can hold a longer, self-checked task — so it's a better place than ever to plan a content run: break a webinar into fifteen angles, batch a week of scripts, or map one long article into a cross-format plan. What it still can't do is make or ship any of it. It renders no media, holds no brand template, and publishes nowhere.
That's the handoff Kompozy is built for, and you can act on it today. Take Grok 4.6's plan or drafts and drop them into Kompozy as sources. Kompozy generates the finished formats each one needs — a HeyGen persona short with burned-in captions, a brand-exact carousel through HyperFrames, quote cards, a blog draft, a newsletter — all in your voice via the Persona Brief, then schedules and autopilots the whole 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 Grok 4.6's output goes in as raw material and the engine ships it. Use the new model to think faster; use Kompozy to turn that thinking into posts your audience actually sees.
Grok 4.6 is an agent-and-coding upgrade at the same 500,000-token context and the same headline price. xAI reports gains across its launch benchmarks (Artificial Analysis Intelligence Index 61 vs 56), better self-verification on long multi-step tasks, a new xhigh reasoning level, and stronger first passes on visual and interactive projects.
Via the xAI API it is $2 per million input tokens, $0.50 cached, and $6 per million output. Once a request crosses 200,000 tokens the whole request bills at double ($4/$12), and a faster variant costs about double the standard rate. It also has 2x included usage in Grok Build and Cursor for the first week after launch.
It can draft the text — scripts, captions, angle batches — quickly, 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.