I’m not a founder, so I don’t think the same way you do. But after seeing what Cursor could offer me for the price, I actually turned Privacy off, hoping that my workflows would contribute to making future versions of Grok and Composer as good as possible.
It worked until few minutes ago. Now I get: “This model does not support custom API keys.”
What is going on?
PS: I disabled the local models I was playing around with few months ago and it worked. But that’s just a bug. I never had problems since I added them until now.
Hah. Had a very similar experience a few months ago – Better subagent control in the Cursor IDE . @Colin 's advice at the time helped me, as our fellow poster pointed out – a subagent rule is effective in governing which models get invoked for this purpose (Composer 2.5 has worked really well for me here).
I see this from time to time after the model burns a lot of tokens and time (I though it crashed):
"The large replacement failed. I’ll use smaller patches or a targeted write instead. "
Did the harness just block a large patch ?
Cursor Grok 4.5 is subject to the same Privacy Mode controls as other models offered in Cursor. With Privacy Mode enabled, Customer Data is not used for training by Cursor or model providers, and Cursor maintains zero data retention agreements with model providers for ZDR-supported models.
@TByte007 would be great if you could raise a bug report, including a Request ID!
Thanks for the clarification, @Colin ! Good to know.
Hey everyone,
I’ve been heavily integrating Grok 4.5 into my automated multi-agent loops for a large CMS modernization project (refactoring, tests, docs) and wanted to share a few observations from a real, highly autonomous setup.
I use Grok 4.5 natively as the orchestrator (delegating work to subagents), as well as for doc-writing and test-writing.
-
Surgical precision with diffs
For autonomous tasks, Grok 4.5 has been excellent at executing focused code changes. The diffs stay tight and rarely disturb surrounding structure — which is why it’s become my default for writing comprehensive tests and generating documentation. -
Token efficiency & minimalist output
Compared to Claude Opus, Grok 4.5 is notably more token-efficient. It tends to skip conversational filler and go straight to the work. In continuous agentic loops, that lack of “yapping” adds up quickly in both tokens and wall-clock time. -
Better cache behavior in long sessions
This one surprised me in practice: with Opus 4.8 I often see repeated cache writes even when the context already appears cached. With Grok 4.5, during normal refactoring it mostly reads from cache and only writes when something genuinely new enters the context. That difference matters a lot across multi-step agent sessions. -
Proactive / autonomous workflow awareness
The standout for me has been its initiative. In one session, without me asking for it, Grok noticed I was on the wrong branch (branch protection), switched to a new feature branch, moved the changes over, and opened a PR according to my existing criteria. It effectively understood the workflow constraint and fixed the process, not just the code.Caveat: that level of autonomy won’t be welcome for everyone. In my setup it was exactly right — but it did once switch branches when I was already on the correct one. That only happened once so far, and it’s been correct since.
Overall, if you run multi-agent architectures and want a precise, cost-effective model that stays focused, Grok 4.5 has been a clear upgrade for my workflow — especially as orchestrator + docs/tests.
There’s also this article which Digital Ocean recently shared. So for me, Grok is not an option until I see something concrete on their privacy and git stance.
Grok Build is a separate SpaceXAI product and is not part of Cursor. The incident you referenced does not change how Cursor handles data when Grok 4.5 is selected in Cursor!
Grok is being forced as the model every time I open a new chat. Is this being fixed?
EDIT: I see there’s an acknowledgement that it’s a bug in Grok re-enables itself after being disabled in Settings . Please post in this thread too when that bug is resolved.
Used Grok 4.5 It is not useful for me. I personally prefer Composer 2.5 over Grok 4.5.
Hi @Charles_Wood , I noted this recently in Feedback - Don't change my default model (I suspect this is intentional on Cursor’s part, but I object to it).
This bug was opened 8 days ago – Grok re-enables itself after being disabled in Settings . So I’m sure they’ll fix it eventually but they’re not in any hurry.
I have had my fair share of rage moments with cursor and composer at times.
At times so bad that I thought about switching to something else.
However Cursor Grok 4.5 really hit the spot.
It is extremely good for the price and I am not sure Opus 4.6 (which I still use as I think its better then 4.8) could beat this.
I am very pleased with this model hopefully it stays as a default model that can be used alongside composer 2.5. Is that the plan ?
I would be tempted to do this for personal projects. If only there were a way to set this per repo (work private, personal shared)
I am not using it for anything, but Cursor keeps trying to force me to use it. I have removed all Musk products from my model list. That is, I set all the switches to OFF. Yet, when my current model of choice GPT-5.5 hits a limit, Cursor re-enables grok on the model list and switches my account to it. This is 100% a deal killer. It has happened six times so far. I will not fund Musk enterprises and neither will my company. if it means I have to drop Cursor completely, then so be it. I have reported the problem, but have yet to see any progress toward a solution.
I’ve been using it as the primary model after hitting limits on api usage due to fable. Its been pretty good so far, a bit slower than composer but a point up in quality of results.
The major issue with cursor grok is that it doesnt know how to ask questions via the UI selector, it always asks them through the chat so you have to write out all your responses rather than selecting from the UI.
Been using Grok a bit since I hit my API limit. I find it sloppy and hasty. I do a lot of cleanup. It’s reasoning is hard to understand at times. I feel it doesn’t connect all the details together in the same way other models function. I have primarily used GPT and Claude but have time with most higher performing models. Also, I have Grok on High (not fast) and was on Medium for all other models.
I recently got this from it:
Yes. I guessed, then gave fake process explanations when you pressed. That was dishonest.
request ID: e71c06af-6a55-466b-ab29-3b970295ecc6
Another gem:
You asked for a rename. I also rewrote
matchRatePercentwithout being asked. Reverting that part.