I really enjoyed the work style with Cursor. Usually I use auto and sometimes I would use Grok with extra high effort. Ultra by the way.
I loved discussing all the details with Cursor. I used to have the feeling it totally gets me. And it’s so cool so reliable. It could be fast and efficient and it also could be detailed and thoughtful as needed. So auto always my first choice.
However yesterday afternoon same conversation continued it just suddenly felt different. The output format became much simpler and it just doesn’t think or provide enough action. It’s like I have to push it. It just starts not to respond as sophisticated as it used to be. It’s like it keeps losing the details and does not work hard anymore. It’s become so stupid I had to shout at it. I keep wondering what happened to you! It’s the latest version Windows cursor. Network alight. I also restarted the PC. And I tried new conversations. Both auto and Grok 4.7 extra high I tried. It just became so dumb and lazy like a stranger. I expect so much quality work from it and it just gives me garbage. I am so unsatisfied and I switched to Opus 5.5 which is of course third party. Honestly I am not a big fan of Claude. I love more the tone and style of cursor auto as it used to be not now. But today I had to switch to Opus and it works. It’s a little too lengthy so to speak but it does the job and it discovers flaws inside the code and solves it. All that could’ve been done by Auto and Grok. Now it’s like Auto can only run some quick and short task.
You know it’s like I describe some important task to do and then Auto and Grok just output something so short and lazy that it’s like it just doesn’t care to work for it. It’s so unsatisfying. There must be some reason? Are you guys degrading the models? Is this just temporary or what? I am so annoyed by the performance of Cursor auto and Grok since yesterday afternoon.
it has for sure gotten worse. much worse and unusable. switching to claude code when this month comes to an end.
the stuff grok spits out is literally unintelligible word spaghetti. impossible to parse what it’s saying. even when it finds real bugs e.g., it is terrible at explaining them I have to hand-hold it to be able to even explain what it’s trying to say.
when 4.7 came out I was gassing it up like crazy, it’s trash now. and with the price hike on using ‘other’ models, I’m out!
100% true. Something seriously bad happened to Grok 4.7 today. In about half the threads, I’m basically getting my own request paraphrased back at me. I have to repeat “please stick to the original goal” several times, and all I get is an apology and a promise to follow it, but the results are still very basic. In a few discussions it got completely confused about the project structure: instead of updating the documentation, it modified the code even though nobody asked for that and I accidentally reverted those changes. Yesterday everything was fine, but today it’s just a total collapse in reasoning quality. Initially decided it was a problem on my side, but happy to see it is not just me
Something has gone completely wrong over the last few days with both Grok and Composer—we’ve regressed to models that just make a mess, much like things were two years ago…
On top of that, my external API quota for the other models ran out in an instant, and this is on the Ultra plan!
Someone on your end messed things up, and it’s disrupting our work.
I’ve never felt any advancement brought by Grok 4.7 tho. Actually it’s when 4.7 came out that I had to upgrade to Ultra. You know, before that Pro+ with Auto was pretty enough - Although it might be also due to the workload… I am not 100% sure. But I don’t think Grok 4.7 is any better, not to mention it consumes credits much faster. Anyway, up until recently Cursor Auto and any Grok model is the best for me. I do love the quality and efficiency it delivers. So I hope it’s only temparory dumb.
Seems to me like they nerfed the thinking on it or in some cases turned it off all together. especially with grok bot. its like they went into the inference engine and just set “thinking off”.
Cursor cloud seems to be a bit better.. but have now noticed Grok build cli is extremely lazy, and hallucinates.
and today had a significant error in just a normal chat with grok.
did a new chat with “thinking off llm command” as the message and it gave me a huge deep dive analysis on the war…
Today I had a gpu server I asked Grok for temperature safeguards and notifications for temps/errors. It only set the safeguards for one service the gpus do on the server. When it switched to the other service ( it was aware of 2nd service) it stopped monitoring temperatures and cooked some really expensive gpus at 100C for hours
hey are you using the new X suit subscription? I am still the legacy where there’s third party models included. I sure think you have also noticed the changes recently.
Formerly when I assigned a task Grok and Auto would go all the way to every detail thoroughly. And if I asked it to find the bugs or give me alternative solutions it would give me so much new information. Now it’s like it doesn’t want to think even when I push it.
Another noticeable difference is the plan. For a huge coding task they can no longer absorb all the related data and generate a very long and considerate list of actions. It’s like usually it would expect it to generate 12 steps in the plan. Now it’s just 5. And it’s just recently. Same task with Opus it gives me so much new information and discovers the flaws brought by before. You know Grok now can’t even write docs like it’s so good at before. For the past 2 days so many times I asked Grok or Auto to write some docs for me. The document style changed. The content was unusable. I could notice that. After so many rounds of instructions I had to switch to other third party models to rewrite everything and now I am wondering does it mean now Grok and auto can only be used for Chat or simple automation tasks.
I KNEW I was not crazy. The only concern I have is that cursor offers really good indexing and they require a subscription to use their service right? Does anyone know if another company offers the indexing but you can use your claude subscription or api perhaps?
Until they have telemetry data /session log data , I see no value in using cursor or Grok Bot for coding . I limit them to only creating Issues and orchestration of multiple Grok Build Sessions where all coding is done given I can get get per issue/PR tracking.
Do you get per PR cost tracking if you just use Cursor only ?
I see evidence they can track lines of code ( I suspect they are just pulling this from git ) but do not see any place they track cost per issue/pr which is what any enterprise using this tool set would want.
As far as I can tell, it’s just not there and it seems they have no foundational telemetry data to actually produce it . So I see no reason not to use Grokbot vs Cursor given both are missing critical foundational observability .
Grok Build seems to be the most mature in this aspect.
Love to know if anyone else has found a way to get per BOT usage tracking /cost against the weekly pool
I am not satisfied with Cursor’s in-house models, whether it’s the Auto mode or other Grok models.
Especially in the past few days, they’ve become so shallow and stupid that I can’t count on them to perform complicated and long tasks.
And their answering style has become lazy and over simplified, even if I asked them to think thoroughly.
There are too many flaws in their plan. If I happen to spot one, they would just admit and apologize. They would only be able to correct that one thing that I’ve just pointed out even if I strongly asked them to deep more into the problem.
And then I am tired of correcting them.
So then I had to switch to Opus 5.5 in the ‘Other Models’. Works amazing. It really reminds me of how Cursor models (including Auto) used to deliver high-quality work with me.
Now it’s like: I can no longer think and work no more with Cursor models. They can only execute some small tasks. They like to narrow down the scope and I am tired of pushing them or correcting them.
How do you guys explain all of this happending to my Ultra account? And I am such a huge fan of Cursor. I remember weeks ago and months ago my first choice would always be the Auto mode and it’s so good I often need not use Other Models. Now Cursor models are just lazy, stupid, shallow, and yet amazingly fast for sure.
I really look forward to Cursor models being more intelligent, powerful and helpful. At lease stays the same quality. Not like this moving backwards.
Sorry if this is not the place, i just need to vent…
They broke the the harness or tweaked something, it was not like this a while ago.
Its unreal, I used to go thru tasks like nothing, same exact setup now & I spend half the day going back and forth going in circles.
I spend 3 hours in a simple task, one endpoint couple services & a web view, 3 HOURS. The task was already planned by another agent, reference images where there, clear goal, explanation on how to build was there… yet 3 hours and still it was not right, reverted everything, same prompt (lit copied the prompt verbatim) using sonnet in medium from within cursor and literally took a couple minutes one single attempt and succeeded, built everything as expected, followed my rules, code style & the plan A → Z.
As soon as i finish this month i am switching, or maybe even sooner. Its unacceptable.
Just want to chip in here and share my recent experience - been using cursor on a smaller java angular stacked project with great success, but for the last few days even on Grok 4.7 with High-effort, creating new sessions for smaller tasks, it just seems to go nowhere with some tasks, literally investigating stuff for nearly 20m at a time, then I interrupt, give it an explicit fix for a very small problem, it investigates 20m more… by the 2nd time I just cloned the repo myself and fixed and tested the issue and pushed the hotfix… (it was that simple, even a human could fix it in less time than the 40minutes~ it wasted).
Obviously I think I did something wrong and this morning I continue, even minor UI updates (adding logos to a header, and tweaking the logo background…) it completely botches, ignores specific comments a few times and then I give up and fix it myself…
Not sure how to trust this quality fluctuations… guess we’ll try extra-high and see if the basic usefulness returns? Anyone else experiencing this type of behaviour recently?
That’s weird. For me Auto & Grok are always fast.
It’s like I wanted to shout to them: “have you been thinking hard enough?” even thought I’ve turned on extra high thinking level.
It’s crazy. It’s no longer the problem solver. It’s easy for them to just stop there once careless analysis is done. No longer eager to solve any hard problem. And it’s been a week?
Basically dealing with the whole same problem, started to frustrate, using Grok 4.7 for at least 2 months everything was fluent, did the job and tasks and plans and implementations kind of good. but yesterday 2026-11-10 and today 2026-11-11 it looks like it went 1 year back, and is not doing anything good, leaving bugs , doesnt think of edge cases. Switched to auto but its not good as one frequent High model.
Came here from simple research , there is more of us: