On 21 September SpaceXAI released Grok 4.7: a new, larger base model, trained for longer, weighted toward tasks that take hours, at the same price as Grok 4.6. In the company's own table Fable 5.1 stays ahead on CursorBench and in the terminal, but it costs several times more.
- Grok 4.7 stands on a new, larger base model and a longer reinforcement learning run weighted toward multi-hour tasks.
- The price stays at $2 per million input tokens and $6 per million output tokens, the same as Grok 4.6.
- In SpaceXAI's own table Fable 5.1 leads on CursorBench 4.0 and Terminal-Bench 4.0, while Grok 4.7 leads on legal tasks.
The table is the first thing worth looking at. Not for the numbers, but because the new model loses in it.
SpaceXAI puts Grok 4.7 next to Grok 4.6, GPT-5.6 Sol and Fable 5.1 and hides nothing. On CursorBench, Fable is ahead. In multi-hour terminal work it is far ahead. Grok 4.7 wins where nobody looks first - legal tasks and electrical engineering.
Where the bill is
Fable 5.1 costs $10 for input and $50 for output. Grok 4.7 costs $2 and $6. Output is more than eight times cheaper, and output is exactly what an agent spends when it works alone all night.
That is the real news. Five and a half points on CursorBench are visible, but they are not a gulf. Twenty points in the terminal are. So if your task is long command-line work, you pay more and take Fable. If it is code, documents and a lot of repetition, the bill starts leaning the other way.
The numbers are theirs, from their own table, measured at high effort levels. What it really costs will be known when someone runs it on a real project for a month and shows the spend.
The 40% discount on Vercel's gateway runs until 27 September. If you are going to try it, this week is cheaper.