Grok by xAI: The Complete Guide to Models, Features, and the API

Grok is the AI assistant built by xAI, Elon Musk's AI company, and it is wired directly into the social network X — which gives it one thing most rivals lack: a live feed of what the world is posting right now. Since September 21, 2026 the flagship is Grok 4.7, a frontier reasoning model with a 500,000-token context window, sitting alongside the cheaper long-context Grok 4.3, a smaller model for coding agents, and separate engines for image, video, and voice.

This guide covers the whole picture: the model lineup, the three ways to access Grok, the features that actually set it apart, how to call it from the API, and where it makes sense to choose Grok over Claude, Gemini, or GPT.

The Grok model lineup

xAI ships a small family rather than one model. The hero above lays out the three you will actually pick between:

• Grok 4.7 — the flagship (launched September 21, 2026). A new, larger base model trained with a longer reinforcement-learning run, a 500K-token context window, four reasoning-effort levels (low to xhigh, default high), and the same $2 / $6 per million tokens as Grok 4.6. See our launch coverage and review.
• Grok 4.6 — the August 2026 flagship, still on the API at the same price. Unless your own tests say otherwise, Grok 4.7 replaces it.
• Grok 4.3 — the cheaper option (launched April 30, 2026) at $1.25 / $2.50, with a larger 1M-token context. Reach for it on long documents and high-volume work where the flagship's extra reasoning is wasted.
• grok-build-0.1 — a smaller model aimed at coding agents, at $1 / $2 with a 256K context.

Two more pieces round out the family: Grok Imagine, xAI's image and short-video generator (the 1.5 release improved image-to-video quality and speed), and a built-in voice mode for spoken conversations. One caveat worth knowing: xAI does not publish a knowledge cutoff for Grok 4.7, so for anything current, lean on its live search rather than what it remembers.

Three ways to access Grok

How you use Grok depends on whether you are a reader, a power user, or a developer:

In X, free or with X Premium. Anyone can use Grok inside the X app, with a free tier capped at roughly ten prompts every couple of hours; X Premium ($8/month) raises those limits. This is the path most casual users take, and it is where Grok's real-time-X superpower is most obvious.

SuperGrok, the standalone subscription. SuperGrok ($30/month, or $300/year) unlocks the full assistant at grok.com and in the mobile apps: DeepSearch, the "Big Brain" thinking mode, voice mode, and generous image and video generation. Heavy users and professionals can step up to SuperGrok Heavy ($300/month) for the highest usage limits, a larger team of agents working on each answer, X Premium+ at no extra cost, and early access to new models; a cheaper SuperGrok Lite tier sits below SuperGrok. (Tiers and prices change often — check x.ai/grok for the current lineup.)

The API, for builders. Developers sign up at console.x.ai, add credits, and call Grok like any other model — more on that below.

The features that set Grok apart

Strip away the branding and four things genuinely distinguish Grok from the competition:

• Real-time X access. DeepSearch can read the live X firehose and the open web, so Grok is unusually good at "what are people saying about this right now" questions that stump models working from a static training set.
• DeepSearch and Big Brain mode. DeepSearch runs an agentic research loop across sources; Big Brain (Think) mode spends extra compute on a single hard problem before answering.
• Voice and image generation in one place. A natural voice mode and Grok Imagine put conversation, pictures, and short video behind a single subscription.
• A looser persona. Grok is deliberately tuned to be more candid and less hedging than most assistants — a feature to some users and a risk to others, since fewer guardrails can mean more confidently wrong or edgy answers.

Using the Grok API

The API is intentionally easy to adopt: it is compatible with both the OpenAI and Anthropic SDKs, so in most cases you keep your existing client and change two things — the base URL and the model name.

curl https://api.x.ai/v1/chat/completions \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.7",
    "messages": [{"role": "user", "content": "Summarize today’s AI news from X."}]
  }'

Pricing as of September 2026, from xAI's model list: Grok 4.7 costs $2 per million input tokens and $6 per million output, with cached input at $0.50. A request whose prompt reaches 200K tokens is billed at $4 / $12 for all of its tokens, so long-context work costs double. Grok 4.3 is cheaper at $1.25 / $2.50 (cached input $0.20), and grok-build-0.1 cheaper still at $1 / $2. xAI also offers sizeable free developer credits through its data-sharing program. Because the endpoint mirrors the OpenAI schema, streaming, tool calls, and structured JSON output all work the way you expect.

When to choose Grok

Grok is the obvious pick when recency is the job: monitoring a breaking story, tracking sentiment on X, or research that has to reflect the last few hours rather than last year. Its agentic tool-calling and large context also make it a capable general reasoner and coding assistant. Where it is weaker is the polish and ecosystem maturity of the longer-established labs, and the looser persona means it pays to verify anything high-stakes. In our own one-attempt build tests, Grok 4.7 finished behind the Claude and OpenAI models in both the three-brief build-off and the Hill Climb round, with the same fault each time: the mechanism was right, and the layer you see was not joined to it. On visual or interactive builds, look at the result before you ship it. If your work is deeply tied to X or needs live information, Grok is hard to beat; for the most rigorous coding or enterprise governance, it is worth comparing against Claude and Gemini before committing.