Guys — you are sending out emails claiming that “independent benchmarks show Cursor Grok performs on par with Claude Opus and GPT Terra.”
Based on intensive real-world use: no way.
I’m working on a highly complex production project and generate up to $2,000/month in Cursor/model costs, so this is not based on a few test prompts.
I actively try(d) to use Grok as much as possible, especially because it is basically included with Ultra. But I constantly have to switch back to Opus to fix errors, clean up implementations and repair architectural decisions.
My experience is simple:
Grok is an advanced junior developer. Opus is a senior software architect.
For clearly defined coding tasks, Grok can be useful. For complex architectures, clean frontend/backend separation, larger refactorings and conceptual work, Opus is still in a completely different league.
So presenting Grok as “on par with Opus” creates the wrong expectation - so make your homework to get on par.
**Generating code is not the same as understanding architecture and complex development. And a cheaper model is not cheaper if Opus has to clean up the result afterwards.
…** I’m writing this while Opus is literally working to repair what Grok has just broken again in my project.
Hi @Gilbert_Kreuzthaler Thanks for your feedback! Grok 4.6 is the most capable model we’ve released yet, but we understand that there is more work to do. 4.7 is going to be even better.
One thing you can do that is really helpful is to provide feedback on the agent’s response when you’re dissatisfied with a particular agent action or response. You can also share particular details about what went wrong. This is a great way to provide direct and actionable feedback because it associates it with that particular request ID.
As always, it’s a good idea to start a fresh chat, watch your context window, and consider using plan mode to map out changes before making them. If you have a particular example of what the agent did well, you can also use that as an example to guide the behavior of the next change that you would like the agent to do.
One last thing to mention, which you might not know, is that Grok 4.6 includes a new XHigh reasoning level that was not available in 4.5.
I fully understand that developing and improving a model takes time — that’s not my criticism.
I’m only responding because of the claim “on par with Opus.” Based on my daily experience with both models, that statement is simply irritating and creates expectations that Grok currently does not meet.
But I’m really looking forward to Grok improving to the point where I can replace the considerably more expensive Opus.
Hi @joeybab3 The thumbs up thumbs down feedback option is currently exposed only in the Agents window. If you upgrade to the latest version of Cursor and you still don’t see it the Agents Window, lmk, and I can take a closer look.
Grok changed a key calculation in a code snippet that had been in place for months, which caused the application to malfunction completely - even though that wasn’t what was intended.
Because of the daily Redis caching, this slipped through all the tests and wasn’t noticed until today.
I specifically had Grok pull up the commit revision for that particular change again and received confirmation that it wasn’t requested.
-–
No. The analysis only says “one plan, all tabs read it” and “Slots/maxSlots” - not “10 total instead of 10 per tab.”
-–
I’ve defined hard rules stipulating that changes not in accordance with the requests/analysis must be confirmed again, and I asked Grok whether hard rules are pointless.
Here’s the answer:
-–
No. They’re included in the prompt and often take effect. They’re not pointless.
They aren’t as reliable as they are with Opus—yesterday I managed to bypass a hard rule anyway. You can’t rely on them.
-–
I actually use Grok quite a bit, and I always have Opus take a look at the Grok code - it works pretty well, and as a mid-level programmer, I’m actually satisfied with Grok.
But this is also in contrast to Opus and my last post here. Opus follows these hard rules absolutely precisely, and you can rely on them being adhered to.
That will have to be a task for Grok 4.7.
