Grok 4.7 by xAI: specs, price, benchmarks and when to pick it

What xAI's Grok 4.7 is, its price, context window and published benchmarks. Data as of November 15, 2026, and when to pick it for coding and images.

info Quick Definition
Grok 4.7 is a language model from xAI released on September 21, 2026, aimed at coding and knowledge work. You use it through the API (grok-4.7), in Cursor and in Grok Build, with a 500K-token context window and text and image input.

What is Grok 4.7 and when was it released?

Grok is xAI’s model family. Grok 4.7 was announced on September 21, 2026. xAI says the model works longer on difficult tasks and checks its own work more carefully, and it offers it through the xAI API, Cursor, Grok Build, cloud platforms and third-party harnesses.

The documentation lists reasoning effort levels: low, medium, high (the default) and xhigh. That means you decide how much the model “thinks” before answering; more effort can mean more output tokens and more time. For the general concept, see reasoning models.

What is it good for?

xAI’s documentation lists it as a model for code and chat, and xAI publishes benchmarks for software engineering, multi-hour office work, law, health and electrical engineering. For a designer or a junior doing vibe coding, the first one is what matters: multi-step coding tasks such as changing an interface across several files, fixing bugs or writing tests, either from Cursor or from a script using the API.

It accepts images as input, so you can pass a screenshot of a screen and ask it to implement it or explain what is wrong. The answer is always text.

What benchmarks did it publish and how do you read them?

xAI publishes these figures in the model announcement (the table uses the xhigh level, except DeepSWE, which a footnote marks as high effort). As of November 15, 2026:

BenchmarkGrok 4.7 scorePublished by
CursorBench 4.0 (software engineering)46.3%xAI
DeepSWE v1.1 (software engineering, high effort)71.0%xAI
Terminal-Bench 4.0 (multi-hour terminal work)37.6%xAI
Harvey Legal Agent Benchmark (legal work)19.6%xAI
EEBench (electrical engineering)64.0%xAI
HealthBench Professional (clinical reasoning)56.7%xAI

How to read them:

  • Each number answers a different question. xAI describes CursorBench 4.0 as a test of longer-running coding tasks. Terminal-Bench 4.0 is presented as multi-hour terminal work. For the rest, the source only gives the category.
  • A low score is not a model failure. A 19.6% on a hard legal benchmark says the test is demanding, not that the model is “bad”.
  • Do not build rankings across labs. Each lab publishes with its own setup, and the same benchmark name may not mean the same thing. To go deeper on this, read AI benchmarks.

How much does it cost and what context does it have?

Data as of November 15, 2026, according to xAI’s documentation:

ItemValue
API IDgrok-4.7
Release dateSeptember 21, 2026
Context window500K tokens
Input / outputText and image / text only
Price with prompt under 200K2 USD input, 0.50 USD cached input, 6 USD output (per million tokens)
Price with prompt of 200K or more4 USD input, 1 USD cached input, 12 USD output (per million tokens)
Knowledge cutoffMay 2026
AccessxAI API, Cursor, Grok Build; xAI does not announce downloadable weights

Two details that change the bill. First, once you cross the 200K-token threshold, all tokens in the request are billed at the higher rate, not just the ones above the threshold. Second, the model has no access to real-time information unless you enable the search tools.

Illustrative example of a calculation: a request with 50,000 input tokens and 5,000 output tokens costs 50,000 x 2 / 1,000,000 = 0.10 USD for input plus 5,000 x 6 / 1,000,000 = 0.03 USD for output, about 0.13 USD. To understand how each piece is counted, see tokens, cost per token and context window.

There is also Grok 4.7 Fast: the same model at 2 times the standard rate (1.5 times with long context), available only in Cursor and Grok Build and not in the public API.

When should you pick it and when not?

Pick it when:

  • You already work in Cursor and want to try a different model for long coding tasks.
  • Your tasks fit in under 200K tokens of prompt and price per token matters to you.
  • You need it to read images (screenshots, mockups) and write code from them.

Do not pick it when:

  • You need downloadable weights to run the model on your own infrastructure. For that, see open vs closed models.
  • Your volume is huge and every fraction of a cent counts. To compare with a model designed for that case, check Claude Haiku 5.5.
  • You want to decide only from one lab’s benchmark table.

Example

Illustrative example: a junior designer maintains a landing page with three components that use hard-coded colors. She opens the repository in Cursor and asks Grok 4.7: “Replace the hard-coded colors in these three components with the color variables from the tokens file and change nothing else.” She reviews the diff before accepting it, runs the site locally and compares the result with an earlier screenshot.

What to try tomorrow: repeat exactly the same request with the model you already use and compare three things: how many extra files it touched, how many tokens it spent, and whether the result looks the same. Two real tasks will teach you more than any table.

Common mistakes

  • Comparing xAI’s 46.3% on CursorBench with another lab’s score as if they were the same measure.
  • Crossing 200K prompt tokens without noticing and paying double for the whole request.
  • Assuming the model knows events after May 2026 without enabling search.
  • Looking for Grok 4.7 Fast in the public API: it is only in Cursor and Grok Build.
  • Accepting code changes without reviewing the diff.

Sources

Frequently asked questions

What is the API ID of Grok 4.7?
In the xAI API the model is called grok-4.7. The documentation describes the naming pattern: no suffix points to the latest stable version and -latest to the latest version; the release notes only list grok-4.7 and mention no -latest alias for this model.
How much does Grok 4.7 cost?
As of November 15, 2026, xAI’s documentation lists 2 dollars per million input tokens and 6 per million output tokens for prompts under 200K tokens. At 200K or more, the whole request is billed at 4 for input and 12 for output. Prices change: check docs.x.ai before budgeting.
Does Grok 4.7 accept images?
Yes in the API: according to xAI’s release notes, it accepts text and image input and answers with text only.
What is Grok 4.7 Fast?
It is the same model with faster responses at 2 times the standard rate (1.5 times for long-context requests). According to xAI it is only available in Cursor and Grok Build, not in the public API.
Can I compare Grok 4.7 benchmarks with other models?
Not directly. Each lab measures with its own harness and settings, so a 46.3% from xAI and a 46.3% from another lab do not mean the same thing. To decide, test your own tasks.