What happened
7 and calls it its most capable model for coding and knowledge work. 6, plus a longer reinforcement learning run on a harder mix of tasks.
The mix leans toward problems that take many hours to finish.
xAI also says the model is better at checking its own work and handling longer context. It was trained to natively understand the Grok Bot harness too, which xAI says helps with conversational tasks and general knowledge work.
The price did not move. Input tokens cost $2 per million, cached tokens $0.50 per million, output tokens $6 per million. xAI says speed matches Grok 4.6 as well.
The docs list a 500,000-token context window, text and image input, function calling, structured outputs and reasoning. Requests beyond the 200K context window are charged at different rates. Batch API is not supported.
Why it matters
Price-performance is the whole pitch. On CursorBench 4.0, a benchmark that stresses longer-running coding tasks, xAI says Grok 4.7 is at the frontier in price-performance.
6 High on that test. 8%.
Those competitors are priced at $4 and $10 for input, and $20 and $50 for output, per million tokens.
So Grok 4.7 does not top that row. The argument is that it gets close to the pack at a fraction of the output cost. Fair pitch, but it is xAI's table, on xAI's page, with xAI choosing the rows.
Note the asterisk too: the DeepSWE v1.1 score of 71.0% carries one in the vendor table. Read the fine print before you rewrite your stack.
A bigger model at an unchanged price is a quiet discount on capability.
What to do with it
- Run your own task.Point Grok 4.7 at one real, multi-hour coding job from your backlog and compare it to what you use now.
- Do the token math.Check your context sizes against the 200K threshold, since pricing changes beyond it.
- Try the effort dial.The docs list low, medium, high and xhigh reasoning efforts, with high as the default, so test where the extra thinking pays off.
- Read the benchmark footnotes.Vendor tables pick their own rows, so treat every score as a starting point, not a verdict.
Ciao, sapiens. The model now double-checks its own homework. Your turn.


