What Grok 4.6 actually shipped

xAI released Grok 4.6 on August 12, 2026. The API name is grok-4.6. The company presents it as a successor to Grok 4.5 aimed at long-running agents, coding, knowledge work, and visual or interactive projects, not a chatbot refresh.

The developer docs list a 500,000-token context window, text and image input, text output, and a knowledge cutoff of February 1, 2026. Reasoning effort is low, medium, high (the default), or xhigh. There is no off switch for reasoning. Tools on that page are function calling, web search, X search, and code execution.

What xAI says changed

The launch note says the supplemental training run was longer than Grok 4.5’s. It used curated model-generated data for reasoning and technical concepts, high-quality engineering data, and an improved optimizer. Grok 4.5 then regenerated SFT traces across reasoning efforts and agent harnesses. Agent RL covered knowledge work, general coding, kernel optimization, web development, and CAD.

xAI’s claim on long jobs is more self-testing. First passes on apps and interactive work are described as more polished than Grok 4.5. Those are the company’s readings of its own evals, not a third-party lab.

The model card dated August 12 (revision August 17) adds that Grok 4.6 is a 1.5T-scale model developed with Cursor, and that it received supplemental training on anonymized Cursor workflow data.

The published score table

The announcement table is xAI’s, mixing developer-reported numbers with public leaderboards. Higher is better on every row below.

The AA Intelligence Index is a composite of nine tests. Grok 4.6 ties GPT-5.6 Sol and sits one point behind Fable 5. On GDPVal-AA it is the highest of those four columns. AA-Briefcase, on the same page, is 1,577 against Grok 4.5’s 1,313.

Coding is mixed. CursorBench is close to Fable 5. DeepSWE and Terminal-Bench trail both named competitors. The honest sentence is not “best coding model.” It is more competitive on IDE-style agent work, and still behind on some terminal and SWE sets.

The model card’s 3DCodeBench row is a separate claim: 54.0% at high effort, above Grok 4.5 at 49.8%, Opus 5 at 49.9%, and Fable 5 at 43.7%. That set grades procedural 3D assets written through APIs, not a person driving Blender.

API price

Base input and output match Grok 4.5. Cached input does not. Once a prompt hits 200,000 tokens, the whole request bills at the long-context row.

Grok 4.5’s short-context cached input is $0.30, not $0.50. Long-context cached input is $0.60, not $1. Workloads that reuse a large fixed prompt will notice that.

A 100,000-input / 20,000-output call, ignoring cache and tools, is about $0.32. The same shape at 250,000 input tokens crosses the 200K line and is about $1.24.

The launch post also names a Fast variant at twice the standard price. The pricing table on the docs site does not give that SKU its own named row. Check the console before you point production traffic at it.

Priority Processing is a separate 2× multiplier for scheduling. Do not mix that with Fast.

Consumer plans

The individual pricing page lists Grok 4.6 as a SuperGrok feature at $30 a month. SuperGrok Plus is $100. The Free plan copy does not name Grok 4.6. Taxes and the billed amount vary by region.

Where you can call it

On launch day: the xAI API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare. Grok Build and Cursor had doubled included usage for the first week.

On August 14, GitHub’s changelog added Grok 4.6 to Copilot Pro, Pro+, Max, Business, and Enterprise. The picker covers VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, JetBrains, Xcode, and Eclipse. Business and Enterprise admins have to turn the policy on. It is off by default. Rollout is gradual.

The useful split is price versus window. Base rates did not rise. The 200K line still doubles the bill. Scores rose on the company’s table, and they did not rise on every coding row.

Source cutoff: xAI launch note, Grok 4.6 docs, API pricing, consumer plans, and model card; GitHub Copilot changelog, as retrieved August 22, 2026.