Today’s Cursor is really bad and completely unusable ！！！
2 Likes
Fully agree. Over-engineering is killing this. I renewed my Pro-subscription today but seriously thinking about moving to windsurf or similar. Issues:
- Composer is gone & now it is agents.
- Chat gets filled up after 8-10 conversations & it recommends new chat.
- You are forced to start new chat beacuse chat simply stop responding.
- When new chat starts, it loses all context & we have to burn requests to bring it back to track
- All of a sudden, it stops working & keeps ‘generating…’ for ever.
- Finally, I dont know from whom these guys are taking feedback from. I love cursor but having said that losing my confidence & seriously thinking about alternatives.
11 Likes
they also remove copilot in cursor, aka we cannot install copilot plugin in cursor
1 Like
Now it does not show Sonnet 3.7 thinking in the models page “Cursor – Models”. It has been replaced with “Max” which is usage based. So the 3.7 thinking model is no longer available for premium users and is not included in 500 fast requests?(Please correct me if i am wrong). If that is the case then this is a deal breaker!
2 Likes
Yeah its shocking how few people actually look for details. With the improvements for thinking and max its not surprising that the extra cost on tokens/requests has to be paid.
1 Like
… wut? How could it possibly have done this?
3 Likes
I just ran into a fun one. I have some prompts I tend to drop into chat to get repeatable behavior in certain situations. Essentially, rules but a bit more manual. This is a snippet of a prompt I use to keep 3.7 from over-enginneering solutions.
- **Document technical discoveries** with `/// SPIKE-FINDING: [what you learned]` comments
- **Highlight unexpected challenges** with `/// SPIKE-RISK: [description]` comments
- **Make non-obvious implementation decisions explicit** with `/// SPIKE-DECISION: [rationale]` comments
Two weeks ago, this was working perfectly. It understood that /// meant the comment in the language it’s writing in. Today, it drops literal /// into the code breaking syntax.
This one is particularly interesting as I haven’t touched this project for 2 weeks. I got pulled into various other tasks. In other words, the only thing that’s changed between 2 weeks ago and today is Cursor.
In this case, the fix is as easy as telling Cursor to be mindful of the language, but it’s a surprising change and a case of something breaking.
2 Likes
lol thats oversimplifying it. No, its not just ‘Cursor’ that has changed. Cursor team has made also improvements on handling of requests, processing, … but also Anthropic made changes on their models (not all were good changes i think).
Sure, but when I use RooCode, I don’t run into the same issues
1 Like
I’ve had enough. Time to look for alternatives, the agent is destroying more than it’s creating.
1 Like
Cursor doesn’t make any money, guys. That’s the problem. They can’t make money selling 500 requests for $20/mo. The economics aren’t there. So they try to make the economics work by eliminating support (sans this forum), compressing your code, and using all sorts of fancy tricks which degrades the quality of the code it outputs. But ultimately, they haven’t yet found a formula that works, as this thread and so many others indicate.
They will either need to increase the price and restore the quality back to where it was in the 0.45 days or it’s going to be a low-end solution.
But I think it’s important to point out that Cursor is subsidized by VC money and is not profitable, and won’t ever be profitable at current token rates because the quality degradation necessary will end up costing them the loss of the customer base.
5 Likes
The quality in the past week is OUTRAGEOUSLY bad. Frequently misunderstands BASIC requests. The hair pulling moment I got just an hour ago was Cursor replying “… I will stop trying to install packages.” when all I said was, “Don’t INSTALL IT THERE”.
WHAT?!?!
Totally sounds like talking to my toddler where his ears are for decoration. Oh. And this is on THINKING mode with 3.7 MAX. I don’t see how that is MAX.
7 Likes
I have subscribed to cursor for 5 months, and it is getting dumber. I used to use Claude 3.5 1022 model, which worked well. But recently, cursor has iterated many versions and added newer advanced models, but the quality of work has become worse and worse. There are more and more code quality and errors. Either you need to fix it repeatedly (which will take up a lot of credits), or you can only fix it manually. It’s hard to describe.
I am constantly switching between multiple tools. I subscribe to Claude ai, anthropic API (using Claude code), windsurf (fast, but credits are consumed quickly) and GitHub copilot pro. I spend at least $80 a month on these places. To be honest, I am very tired now. I recently plan to use Claude code as the main force. It is really great, at least in terms of code context and automatic editing capabilities, which is much better than cursor, because it rarely leaves errors in the code and avoids multiple reworks. Then I plan to use cursor as a deputy, leaving some small functions to cursor to complete. I am also using the preview version of vs code, which supports proxy, but the effect is average at present and the speed is much slower. It depends on whether the subsequent iterations will be better. As for windsuf, I think the credit deduction is not transparent. I spent $15 in 3 days and even added $10 to buy 300 flow credits, but it was gone in less than 1 day. I almost don’t know how it was consumed so quickly. If cursor can improve the code quality, I think I will continue to support them.
6 Likes
Depends how many tokens those 500 requests use. When using Claude Workshop directly I was paying about 0.02$ for each thread (multiple queries). That would be 10$ for 500 threads (multiple requests).
A bulk B2B collab could fetch discounted pricing, not the retail rate.
Then there are people that do no use the whole 500 requests.
And there are people that forget to unsubscribe.
2 Likes
even my chat history was deleted. can one go back to old version 0.45 & still not lose chat history? anyone?
I went back to 4.5.14 and it seems to be okay… the new UI on 4.7 was so bad, not having restore points after responses? Wtf I can’t believe they removed that. That was literally one of the best features.
Things have definitely been getting bad for me as well. It doesn’t look like cursor is utilizing my cursorrules at all, when it was actually doing a great job if previously. Now it will just sit there and freeze, no responses, nothing. But when I look in my portal, it shows that it received the request and generated responses. I even tried using a “premium model” I just end up with $15 of charges and no progress forward. Very frustrating. As to someone saying they can’t make money, I don’t mean to be a jerk, but it’s a companie’s responsibility to figure out their finances, not the end-users. If they need to charge more, than that’s what they have to do, unfortunately. But to build a product that just doesn’t work well, that’s not a great way of going about things.
3 Likes
Try going to the Cursor Settings, there’s a Rules section, and a checkbox in the Project Rules section that says to include .cursorrules files.
I can see they’re still developing the Rules system, and we’re still figuring out how to use it. It’s seeming like in the project’s .cursor/rules/ using .mdc files, with some yaml headers, is the “right” way now. Back in the Settings try creating a Project Rule and try the different sorts of “Always Apply” “Manual” etc. settings, and look at the files it creates in that new rules directory.
JBB
1 Like
I’ve tried all that. Unfortunately, it hasn’t resolved my problem. I could live with the rules not working 100% if the AI would work. I’ve been using Agent and it does nothing right now. Just sits there saying it’s working and nothing else.
1 Like