Grok 4.7 just landed in GitHub Copilot and Cursor — here's what actually changed for you

If you use GitHub Copilot or Cursor, you don't need to do anything to get Grok 4.7 — it's either already in your model picker or arriving there this week. xAI released it on 21 September 2026, and the rollout was unusually fast: live in Cursor and Grok Build on day one, with GitHub switching it on gradually across every paid Copilot plan the same day. That's the practical headline. The more useful one, if you're deciding whether to actually pick it for a job, is what's changed underneath and what it costs you to run.

xAI describes Grok 4.7 as built on a bigger underlying model than Grok 4.6, trained for longer on harder problems — specifically, tasks that take many hours rather than a few minutes to work through. In plain terms: it's meant to stay useful for longer without losing the thread or making a mistake it doesn't notice.

That self-checking improvement matters more than it sounds. If you set an AI agent loose on a coding task that takes an hour of back-and-forth, a model that catches its own mistake at minute ten saves you from a mistake compounding for the next fifty. xAI also trained Grok 4.7 to work natively with its own agent tool, Grok Bot, which it says helps with everyday conversational use too.

The headline price hasn't moved: $2 per million input tokens, $6 per million output, exactly what Grok 4.6 charged. A million tokens is roughly 750,000 words, so for most individual jobs — a code review, a research summary, a first draft — you're talking cents, not pounds.

And that second number moved. Independent testing by Artificial Analysis found Grok 4.7 needs roughly 81,000 tokens, at its most thorough setting, to finish the kind of task it uses to score models — against about 36,000 for Grok 4.6 doing the same. Same meter rate, more than double the tokens spent. If you're on a flat Copilot or Cursor subscription, that's invisible to you; the platform absorbs it. If you're calling the API directly and paying per token, it's the number that decides whether Grok 4.7 is actually cheaper for your job or just cheaper-looking on the rate card.

On xAI's own comparison, Grok 4.7 beats Grok 4.6 on every benchmark the company published — coding, legal reasoning, electrical engineering, multi-hour office work. Against Anthropic's pricier Claude Fable 5.1, it's a mixed picture: Fable still wins on general coding and sustained terminal work, Grok 4.7 wins on electrical-engineering problems and legal-document tasks. Independent testers at Artificial Analysis measured its terminal-command performance notably lower than xAI's own published number, which is worth knowing if your work leans heavily on an AI running commands in a terminal rather than writing and explaining code.

The practical advice, in one line: if you're already paying a flat subscription for Copilot or Cursor, switch the picker to Grok 4.7 and see if you like it — there's no extra cost to trying. If you're paying by the token through the API for high-volume work, don't take the $2/$6 sticker at face value; run your own real job through it and compare what you're actually billed against what Grok 4.6 or another model would have cost for the same finished result.