Grok 4.7 is more interesting than the version number makes it look.
This is not just “Grok, but slightly smarter.”
The real story is that xAI is clearly pushing Grok toward becoming an AI worker: something that can code, operate tools, work inside terminals, reason across long tasks, inspect its own output, and keep going.
And the price makes that especially interesting.
Grok 4.7 starts at $2 per million input tokens and $6 per million output tokens.
That matters because the future of AI is increasingly not:
prompt → answer
It is:
goal
↓
search
↓
read files
↓
write code
↓
run tools
↓
inspect results
↓
fix mistakes
↓
repeat
↓
finish the work
And agents burn tokens.
According to xAI’s published evaluations, Grok 4.7 improves significantly over 4.6 across coding, terminal use, engineering, and long-horizon professional work.
One of the biggest jumps is Terminal-Bench:
Grok 4.6: 20.3%
Grok 4.7: 38.0%
That is close to an 87% relative improvement.
CursorBench also moves from 40.4% → 46.3%.
DeepSWE moves from 65.2% → 71.0%.
And its engineering benchmark jumps from 53.0 → 64.0.
The numbers themselves are not the entire story.
The direction is.
xAI appears to be optimizing Grok less around “give the smartest answer” and more around:
“Give this model a difficult job and let it work.”
Grok also has access to things that become increasingly valuable in an agentic workflow:
• web search
• native X search
• code execution
• file search
• function calling
• remote MCP servers
• document retrieval
• image generation
That makes Grok particularly interesting for developers.
Imagine giving an AI access to your repository, logs, database tooling, cloud APIs and terminal, then asking:
“Find why this production issue happens, reproduce it, patch it, test it, and explain exactly what changed.”
That is where models like Grok 4.7 are heading.
Not better autocomplete.
Not better chat.
Software that can actually do work.
And at $2/$6 per million tokens, xAI may be making one of the most aggressive bets in the frontier-model market:
Maybe the winner will not simply be the smartest model.
Maybe it will be the model that is smart enough, fast enough, tool-capable enough, and cheap enough that you can afford to let it keep working until the job is finished.
That is what makes Grok 4.7 worth paying attention to.
Grok 4.7 is here.
It's a notable improvement over Grok 4.6 at the same price and speed.