Grok 4.7 gives you a 500k-token context window, image input, four selectable reasoning-effort levels and a two-tier price that starts at 2 dollars per million input tokens — that is the headline list of grok 4.7 new features, released by xAI on 21 September 2026 and available on the xAI API under the model name grok-4.7, per xAI's official release notes. It is a flagship aimed squarely at long coding sessions, agent workloads and professional knowledge work, and it lands in a crowded week: Anthropic shipped Claude Opus 5.5 one day later, on 22 September 2026.
📺 Watch: Elon Musk’s NEW Grok 4.7 is WILD! 🤯
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
Grok 4.7 New Features: The Full Breakdown
Working from xAI's official release notes, here is what Grok 4.7 actually ships with:
- 500k context window. Enough room for large codebases, long agent transcripts or a stack of research documents in a single request.
- Text and image inputs, text-only output — and notably, no text output limit on the API.
- Reasoning effort controls. Four levels — low, medium, high (the default) and xhigh — so you decide per request whether speed or depth wins.
- Encrypted reasoning on the Responses API. Grok 4.7 always returns its reasoning as encrypted content regardless of include parameters, which matters if you log responses in automation pipelines.
- US regional endpoint availability for teams with data-residency requirements.
Beyond the spec sheet, launch coverage from 21-22 September 2026 (Crowdfund Insider, Yahoo Tech and others) reports that Grok 4.7 runs on a new base model of roughly 2.1 trillion parameters — about 40 per cent larger than its predecessor's 1.5 trillion — and that xAI supplemented training with internal engineering records from SpaceX, including Starlink satellite telemetry and failure-analysis data, to strengthen hardware and physical-systems reasoning. The same coverage describes a model that stays with demanding assignments for longer and reviews its own output more carefully, with xAI positioning it as the strongest Grok yet on refusal quality and jailbreak resistance.
Grok 4.7 Pricing: The Two-Tier Structure
Grok 4.7 pricing splits on prompt size, per xAI's release notes:
| Prompt size | Input | Cached input | Output |
|---|---|---|---|
| Below 200k prompt tokens | 2 dollars per 1M | 0.50 dollars per 1M | 6 dollars per 1M |
| Above 200k prompt tokens | 4 dollars per 1M | 1 dollar per 1M | 12 dollars per 1M |
Two pricing details deserve attention. First, the tier doubles once your prompt crosses 200k tokens, so a workflow that habitually stuffs the full 500k window pays twice the rate of one that keeps prompts lean — worth designing around in any high-volume Grok 4.7 automation. Second, there is a separate variant: Grok 4.7 Fast, the same model served at twice the token rates, available only through Cursor and Grok Build rather than the public xAI API. For comparison, Anthropic's Claude Opus 5.5 announcement of 22 September 2026 lists 4 dollars input and 20 dollars output per million tokens — so Grok 4.7's sub-200k tier undercuts it on paper, while real value depends entirely on how each model performs on your workload.
If you want to turn model releases like Grok 4.7 into actual income — with agent workflows, prompt libraries and five live coaching calls a week showing you what to build — check out the AI Profit Boardroom → get inside here. Want AI SEO help 1-on-1? Book a free SEO strategy session and get a plan for your site.
How Grok 4.7 Fits an Agent Stack
The specification reads like it was written for agent harnesses: a big context window for long-running sessions, reasoning-effort control so a router can dial depth up or down per task, image input for screenshot-driven workflows, and no output cap for long generated artefacts. If you run an agent framework, Grok 4.7 slots in as a brain alongside the models you already route between — the Hermes bot vs Grok bot comparison covers how the two ecosystems differ in practice, and the Goldie Bench write-up covers how these model brains compare in hands-on tests.
The competitive context makes Grok 4.7's positioning clearer. Within roughly a fortnight the market saw DeepSeek push its V4.1 architecture, Anthropic ship Claude Opus 5.5, and xAI answer with Grok 4.7 — three flagships aimed at the same coding-and-agents money. If you are benchmarking options for your own stack, the Kimi 2.6 benchmark breakdown and the Hermes with GPT-6 guide show how the other frontier brains have been landing, while the DeepSeek harness guide and DeepSeek V4 on Ollama walkthrough cover the budget end of the same decision. The pattern across all of them: pick the model per task, not per loyalty.
📺 Watch: Grok 4.7 + Hermes Agent is SCARY GOOD!
What Grok 4.7 Does Not Change
A few boundaries keep expectations honest. Grok 4.7 outputs text only — image input is supported, image generation is not part of this model, and there is no audio modality on the API per xAI's release notes. The Fast variant is fenced off: if your tooling is not Cursor or Grok Build, the double-rate low-latency serving simply is not available to you, and the public API tier is what you get. The always-encrypted reasoning content on the Responses API is a real constraint for teams that audit chain-of-thought in their logs — you receive reasoning, but not in readable form. And the xhigh reasoning-effort level, while useful for the hardest problems, burns more output tokens at 6 or 12 dollars per million, so leaving it switched on globally is the easiest way to quietly double a bill. None of these is disqualifying; all of them belong in your routing logic before Grok 4.7 takes production traffic. If you already route between a hosted frontier model and a local fallback, the DeepSeek harness vs Hermes comparison shows what that decision framework looks like applied to a different pair of options.
What Grok 4.7 Means for Money-Making Workloads
For anyone building revenue on top of AI — SEO content systems, client automation, coding services — Grok 4.7's economics are the story. The sub-200k tier at 2 dollars input and 6 dollars output prices a frontier-class model at commodity rates, and the cached-input rate of 0.50 dollars per million tokens rewards exactly the repeated-context pattern that agent systems produce. The discipline it demands is prompt hygiene: cross 200k tokens and every rate doubles. Structure your agents to summarise and reset context rather than endlessly append, and Grok 4.7 stays in its cheap tier; let transcripts sprawl and your margin halves. If you want a structured way to build that discipline into your setup from day one, the Agent OS guide covers the operating-system approach of giving agents structured instructions and workflows that keep sessions lean.
Grok 4.7 New Features FAQ
When was Grok 4.7 released?
xAI released Grok 4.7 on 21 September 2026, per the launch coverage and xAI's September 2026 release notes. It is live on the xAI API as grok-4.7.
What is the Grok 4.7 context window?
500k tokens, with text and image inputs and text-only output. xAI's release notes also state there is no text output limit.
How much does Grok 4.7 cost?
Below 200k prompt tokens: 2 dollars input, 0.50 dollars cached input and 6 dollars output per million tokens. Above 200k prompt tokens those rates double to 4, 1 and 12 dollars respectively, per xAI's release notes.
What is Grok 4.7 Fast?
The same model served at twice the token rates for lower latency, available only through Cursor and Grok Build — not on the public xAI API, per xAI's release notes.
Can Grok 4.7 control reasoning depth?
Yes — reasoning effort supports low, medium, high and xhigh levels, with high as the default. Dial it down for routine tasks to save output tokens, and reserve xhigh for genuinely hard problems where extra thinking pays for itself.
Verdict on Grok 4.7
Grok 4.7 is a serious spec at an aggressive price: frontier-scale parameters, a 500k window, controllable reasoning depth and a sub-200k tier that undercuts Claude Opus 5.5 on paper. The open question is always the same one — how it performs on your workload rather than in launch material — so treat the claims as sourced above, run your own comparison on a real task, and let the results pick your default brain. Whichever way that lands, the direction of travel is unambiguous: frontier capability keeps getting cheaper, and the people who win are the ones with systems ready to use it.
If you want working AI systems that turn cheap frontier models into revenue — the full Agent OS, 1,000+ workflows and weekly live coaching — check out the AI Profit Boardroom → join here. Prefer a personalised plan first? Book a free SEO strategy session and map out your next 90 days.
