Now that Grok 4.7 is available, we’d love to hear how it’s working for you!
Some things we’re especially curious about:
How does it handle long, difficult tasks? Grok 4.7 was trained on a harder mix of problems weighted toward work that takes many hours, and it’s meant to verify its own work more carefully and manage longer context better than Grok 4.6. Does that hold up in your sessions?
How is it on non-code knowledge work? This release was trained to natively understand the agent harness, so it should be noticeably better at conversational tasks, research, and producing documents and presentations. If you’ve used it for anything beyond code, we want to hear how it went.
How is it on visual and interactive projects? If you’ve pointed it at UI, apps, or other interactive work, tell us how the first pass compared to Grok 4.6.
How does it compare to Grok 4.6, Composer, and other frontier models in your day-to-day?
Drop your thoughts below (and leave a Request ID if you have one)! We read everything.
IMHO: Grok 4.7 is a sidegrade. Yeah, it is a bit smarter, but per task, it spends almost twice as many tokens as Grok 4.6 to do so. (Referring to AAII, so YMMV) It is no longer that kind of value frontier like Grok 4.5. What a run that was. This time quite a letdown tbh.
SpaceXAI did not compare to Astra; Cursor even disabled Luna and Spark in the default view in CursorBench, I wonder what that means.
Come on guys… I want to cheer for you too. Not like this…
Plans on half price or resetting quota? Or unless you’ve got Composer 3, please…?
I am trying grok 4.7 on a live application should see what will be the result? I am also using Grok 4.7 High only not Fast? I’ll share the result after few hours.
I have a question about the model naming convention. It’s kind of confusing to me. Why is Grok 4.6 called “Cursor Grok 4.6,” while Grok 4.7 is now just “Grok 4.7”? Is Cursor going to release a fine-tuned model specifically for coding, or is this just misleading naming?
I’m not sure Extra High is really necessary, since the High mode in 4.6 didn’t show a significant improvement in performance. In practice, I still had to add extra prompts to guide it toward the right direction. That said, I’ll test Extra High today to see if it can pinpoint issues more accurately. I’m looking forward to seeing better performance from it.
I’ll be sure to give it a try. I’ve been using 4.6 Extra High Fast on multiple projects simultaneously and still only have hit 32% usage this cycle (resets in 14 days). I’ll pay attention to token burn rate for 4.7. Hard to believe it would be significantly better for backend/pure code work, but maybe for documentation and prompt-writing, it will prove to be superior.
Hi @AndreDev, Thanks for your question! Both models are part of the same family, and Grok 4.7 is an evolution and our latest model! The renaming isn’t meant to reflect any differences in the model, rather just simplifying and rallying around the Grok name.
Hi @IDKHTCXD, Thanks for your feedback! Have you experimented with Grok 4.7 on Low or Medium effort? You might find that it is quite capable at a better value relative to tokens spent. We don’t have immediate plans for a launch promotion but we’ll keep you updated here if that changes! We are frequently resetting Grok Bot quotas, the latest one was on Friday for anyone who attended the Grok Bot stream. https://x.ai/galaxy
Oof, it’s super spendy for what I was able to get out of it. I’m not sure the value is there.
I really hope you haven’t abandoned Composer. That has become my daily driver but it’s feeling like an unloved stepchild.
Hi @PizzaConsole Thank you for your post. This is not meant to be misleading! Fable 5.1’s highest available reasoning level is Max, and we’ve compared it to Grok 4.7’s highest available reasoning level, XHigh. In our previous Grok 4.6 release, we did not benchmark XHigh. It is quite costly to benchmark a model on Cursor so that’s why sometimes not every single model or every variant of model is benchmarked!
I can’t comment on its coding ability or token usage just yet, but almost more important than those two things for me is that it is far easier to communicate with. Grok 4.6 was absolutely unusable for any kind of in-depth explanation or architecture design work unless you wanted to get into a whole bunch of jargon and figuring out the terms Grok had made up with no context. I was constantly having Gemini 3.8 translate until Cognition came out with SWE-2, and I moved over to that.
So far, however, it looks like Grok 4.7 is far easier to communicate with.
I 100% agree, however, that the effective loss of the composer series (for now…?) is severely felt.
Whether it was intended to be misleading isn’t really the issue. The problem is that the chart invites a direct 4.7 vs. 4.6 comparison while benchmarking them at different reasoning levels.
That makes the generational comparison inherently confounded. You can’t tell how much of the improvement comes from Grok 4.7 itself versus simply running 4.7 at XHigh instead of High.
If Grok 4.6 XHigh wasn’t benchmarked, then Grok 4.7 High should have been included so there was at least a like-for-like High vs. High comparison. You can still show XHigh as 4.7’s maximum-performance result, but it shouldn’t substitute for the comparable benchmark.
The pricing row makes this even more problematic. Showing both 4.6 and 4.7 at the same $2/$6 per million token price directly above the results further implies an apples-to-apples performance/cost comparison. But XHigh uses a higher reasoning budget, so the same per-token price does not mean the same cost per benchmark task. More reasoning tokens means more actual spend to achieve that score.
So we’re simultaneously changing the model generation and the reasoning/computation level, while presenting the results in a table that visually reads as a straightforward 4.6 → 4.7 improvement. That’s exactly why I consider the presentation misleading, regardless of intent.
Grok Bot doesn’t have a model picker, and we don’t publish which model it’s serving. Cursor manages that, and it can change over time. Grok 4.7 is available in Cursor today if you’d like to try it there!
So far not impressed with Grok 4.7 because its picking up all the bad traits of the Anthropic models. Becoming overly verbose, imprecise, and presumptuous about you knowing what its actually talking about because it had it in a tool use etc. that you don’t readily see.
Such as a review, its saying what the current state is but does not differentiate if its a finding of an error or confirmation it passed! Or it says so using too much verbosity in the introduction and no “Findings High/Medium/Low” like it used to, or bullet points to differentiate one finding from the next.
Its also picking up all the LLM prose annoyances, forced cadences, over use of key words like “holds” instead of “confirmed”.
Grok 4.5 was a joy to work with because its output was so concise and well focused - and thats totally gone.
) It is no longer that kind of value frontier like Grok 4.5. What a run that was. This time quite a letdown tbh. 
