How good is Grok 4.7 at coding?

Grok 4.7 came out on September 21, 2026, from SpaceXAI, the company formerly known as xAI. The company presents it as its strongest model yet for writing code and for office-style knowledge work. It is built on a new, larger base model than Grok 4.6 and was trained longer on hard tasks that take hours, with more practice at checking its own work. It kept Grok 4.6's price: $2 for each million tokens sent in and $6 for each million it writes, where a token is a small chunk of text, about three-quarters of a word.

You can use it in Cursor, in Grok Build (SpaceXAI's own coding agent), through the API as grok-4.7, and through several third-party coding tools. It reads up to 500,000 tokens at once, and its knowledge stops in May 2026.

The company's numbers

SpaceXAI's launch table sets Grok 4.7 beside its predecessor, an older OpenAI model (GPT-5.6 Sol) and Claude Fable 5.1. Three coding rows:

*At high effort; the other Grok 4.7 figures use the higher "xhigh" setting.

Read carefully, the table is honest about the gaps. Grok 4.7 is a big step up from 4.6, roughly level with Fable 5.1 on DeepSWE, but well behind it on terminal work. Note also that the comparison set leaves out the newest rivals, such as Claude Opus 5.5 and GPT-6.1 Sol.

The independent numbers

Artificial Analysis, a company that runs the same tests on many models itself, tells a cooler story. It gives Grok 4.7 at xhigh effort a score of 34 on its Intelligence Index, a blend of nine tests including Terminal-Bench 4.0, against a median of 41 for comparable models. In its own Terminal-Bench 4.0 runs, Grok 4.7 scores about 25.8%, where Claude Opus 5.5 scores about 59.6%, Gemini 4 Argon about 57.1% and GPT-6.1 Sol about 56.1%.

It also flags Grok 4.7 as very wordy: running the index took it 140 million output tokens, against a median of 63 million. That matters for your bill.

A worked cost example

Low prices per token can be offset by high token use. Suppose a coding task needs 100,000 output tokens on a typical model, and Grok uses about twice as many, which matches its ratio on the index above. Ignoring input for simplicity:

The figures are illustrative, but the lesson holds: compare cost per finished task, not price per token. Artificial Analysis puts Grok 4.7's average cost per index task at $7.51.

A good first task is one with a clear finish line:

In this project, the function parse_date in utils/dates.py fails on dates written as "2026-10-28T09:30Z". Add a test for that format, fix the function, run all tests, and summarize the change in two sentences.

Why do a lab's own scores and independent scores differ?

The Terminal-Bench gap, 37.6% from SpaceXAI against about 25.8% from Artificial Analysis, is large, but it is not a sign that anyone is lying. Several ordinary choices move a score.

The agent around the model. A coding model always runs inside a harness, the software that hands it files, runs its commands and decides when the task is over. A lab can tune its own harness for its model; an independent tester usually runs every model through one standard setup. Settings matter too: effort level, time limits, how many retries are allowed, and how many tokens a model may spend. A lab reports the best setting it found; a tester may pick one setting for all.

Which tests, and which tasks. Labs choose the tests where their model shines, and small task sets swing by several points from run to run. Even rival labs disagree: SpaceXAI's table lists Fable 5.1 at 57.9% on Terminal-Bench 4.0, while Anthropic reports 55.8% for the same model.

So treat a lab's numbers as the ceiling under good conditions, and independent numbers as a closer guide to what you will see with standard tools. Then trust your own small test most, because it measures the only workload that matters to you.