Yea i usually back them but i think the rush to bring in Agents because of windsurf is probably costing them more than they anticipated is my guess. Agents iterate whereas the previous composer was turn based. I love agents, but i have slowly started to see a reduction in context window too making it difficult to work with in complex situations so my guess is they are struggling. One of the reasons i prefer cursor is because you could use slow request after paying your flat fee.I don’t want to have to keep topping up no fast request otherwise i had rather go with alternatives. I don’t mind the slow requests, just don’t time out when i have been allocated my turn. However, as i said, i can see this being a problem in Agentic mode because of the iteration, determinism and therefore you could get pushed to the back of the queue during an iteration, hence time outs. Personally, i think unless companies like cursor have their own models that can rival the claude’s etc its a tough business to be in . All speculation and my personal opinion btw.
I think it’s a valid point. Windsurf had to turn around as well, it went from unlimited to 500/month real fast.
I use the chat most of the time and I don’t see any queue even after using all the 500 fast request. Not sure if it’s a bug. In the composer you get put in queue.
If they could upgrade the VS code version Cursor is build on, so we can install the latest Github Copilot (with Sonnet) and offload some of the queries there. There is a lot of chat request that I could make there instead.
Yes but that will break their business model a bit if they did. Ideally, they had rather you use the model through them in a shared pool because not everyone will be coding at the same time meaning they can still make money charging the flat fee. Problem with that is you hope you have enough resource or people have a downtime. However, because of the popularity we are getting an influx of all sorts. i wont be surprised if even non-developers subscribe to cursor instead of say claude web because you get multiple models, have web search etc. Their business model is fine with the slow request , just need to expand to cater for growth .
I think the point is, no one is wanting to purchase extra requests only to be using them to fix the errors that the composer/agent creates. So we are trying to get our ‘500/1000 requests’ out of the slow requests so it feels like we actually didn’t get ‘scammed’ out of what we are paying for.
It feels like we were at first given a full agent, then we were dropped back and the agent started truncating stuff, not saving full responses & was doing cost cutting/reducing context things which have been causing the degradation of the agent/service.
I’d be down for a hybrid approach, hosting local version for calls that are just querying data, logging details etc then using the claude etc for the proper coding requests.
As I’m guessing right now is it having to make multiple ‘slow’ calls which is causing the huge backlog and if one ‘request’ needs to make 2-c slow calls and it might work once, fail then nothing happens? Seems a waste.
Or offload some simple tasks off to the non-premium models.
And when you say its handled off to non-premium models when the premium are full? Are we notified that it has happened or just business as usual?
I don’t want to cost a company funds by expecting my $20/40 to get me unlimited amounts, but I do expect to get the $20/40 value (which is working requests, not me shouting at the agent for deleting things again) ha
Well yes you are right. Thing with all this is we are on unchartered territory . What Cursor is doing is very good but its only a matter of time before the competition catches up.They are early movers so have had to suck up some of the cost and early days with lower user base and it was all good. However, quality of code outputs consistently is what dev’s want but that cost them a lot of money i would imagine. Unless model’s get better and cheaper its going to be difficult to maintain a fair price model is my guess. Alternative, is to have their own base model that can complete with the claude’s etc. However, as a dev if claude released a better model you would want to use that model right?Currently, i don’t even use openai 4o because of the code quality . Alternative, is have a hybrid approach as you mentioned but that will create a lot of inconsistencies i can see. One of the reasons they probably went for a full ide instead of extension is to guarantee these consistencies with models , tab completion, composor, agents etc. Have everything working together and introduce new features.For this you need models that can guarantee quality outputs. Another option is to allow people to use their own open source models , again that will break that consistency and quality of output for the product.Also would you be willing to pay monthly subscription to only use your own model and have constant issue’s because code completion breaks and other parts breaks? Too much headache for them. With time this will get better , but other competition will emerge. For now i just wish they find a solution that just works .
I have the business plan. my mail: onlyvisionsoftware@gmail.com and I also get this. And there is no option to buy more anywhere?? please help me cursor ai support.
It may be true that cursor doesn’t secretly reroute users to lesser models. (I can’t verify; I can’t see under the hood.)
But, stating that as if it settles anything seems disingenuous. As it stands, it might as well.
Today, in particular, it was devastatingly clear that Claude (via Cursor) is being starved of data.
Today, in chat, even when i am attaching a lone file (of about 600 lines), and saying: look in here for the function that is causing the error, it says: I see no such function.
Just to be clear: we’ve been discussing this function by name. We have gone back and forth talking about this function. I then supplied this 600-line file and told it to look at the function. Laziness on my part? I suppose. I could have found it and pasted just that part. But part of the appeal of AI is not having to dig for every little thing, when the AI should be able to find it.
In response, cursor/claude says that the function is not present.
I then copy and paste the 16 lines that comprise the function, and say: did you not see this, when you looked at the file? It says that it did not.
This always happens, on and off. The AI becomes maddening - worse than useless - and I feel as if I’m losing my mind. But eventually it becomes clear that the reason the it’s so stupid is, it’s not seeing the things I send it.
But I have never seen such extremes before today. Today, composer wasn’t working at all (for obvious reasons), but even chat was flat-out not getting it. (The incident I mentioned was from Chat, btw, not Composer.)
If this is what Cursor will be, I deeply regret subscribing for a year. (In retrospect, I was foolish.)
If slow requests are slow, so be it. But they’re not just slow; they’re often much, much worse than worthless. You’re constantly, dynamically changing the product, and you’re doing so invisibly, right beneath our feet. At any given point in time, we have no idea if we will have the good, useful Claude that has been fed the necessary data, or the evasive, worthless, information-starved Claude that stalls, makes excuses, and suggests horrible, destructive code.
You have given us zero ability to gauge what the current state of the cursor/claude relationship is at any given point in time, with any given request. We can only find out through time-consuming, painful trial-and-error what we are dealing with. And we have to be ever-vigilant, because even if it’s going well, it can turn on a dime.
Maybe this model isn’t working for Cursor, economically.
Maybe you should have thought of this before promising things you can’t (or won’t) deliver.
I’m having the same issue. Somehow this affects only Composer, not normal Chat.
What happen to the old slow request system? - this is crap compared to what has been.
I get that ‘slow requests’ means I might need to wait a bit longer for a response, but it’s been completely unusable for me. I keep getting error messages, and even when I finally get a response, it cuts out and I see the same error message again.
This isn’t what you promised when I signed up! I trusted you guys and even bought a whole year’s subscription. Please fix this and make sure the Composer works as you advertised.
Or you could just make Cursor have a setting to use Ollama.
It’s clear that Cursor has oversold its capacity. Promising fast requests and premium model access, only to require additional payments to bypass the “slow queue,” is misleading. Users should not be forced to pay extra to access functionality they were led to believe was already included in their subscription.
The performance issues are unacceptable. Timeouts, degraded outputs, and unusable slow requests indicate that the system is overstretched. When models are swapped or downgraded without warning, it fundamentally undermines trust in the service.
This is not what users signed up for, and the constant pushing of additional costs makes the situation worse. Cursor needs to take responsibility for these failures and stop placing the burden on paying customers to cover their inability to meet demand.
Agreed.
Same also on business plan and right after I set the limit for 100$ for fast requests (it was turned off previously), I still get the “using slow queue” message (even after restarting Cursor). So probably also need a better way of identifying the change of limits.
It would go a long way to hear from the Cursor team about what methods they’ll provide to alleviate the strain on the system. Alternate solutions aren’t really acceptable from a consumer trust perspective, but I think some users may be willing to put up with the situation if given an out (such as well-supported 1st party guides and walkthroughs for local models).
My thoughts on the cost controlling measures (slow queue + degraded agent performance) are:
- The slow queue is problematic because it fails outright to get the job done, not because it’s slow.
- The degraded models are problematic because they’re vastly inferior. They don’t solve the problem.
At the EOD I would trade speed for usefulness.
…because it is unlimited, we have to limit it…
I was trying to give you guys the benefit of the doubt, but you pretty much just made it clear, you guys had a good idea, but then you go a pull this bait and switch move that is reserved for corporate america.
I think everyone will agree that the terminology used in the pricing page implies that we would be able to continue using cursor after our “500 fast requests” we would be able to continue using the service, thats what unlimited means…
but instead we get sandbagged immiediately after the 500 fast requests and have to come to the forum and dig around to find out that were tricked because although we get unlimited slow requests, cursor never promised any responses to slow requests…
seriously, do you really think your userbase is that dumb?
windsurf doesnt even have to beat you when you do this type of shit.
just be honest. dont play word games when it comes to pricing.
Yeah, now this “Unlimited slow premium requests” on the price page is like a lie. At least developers could try to support it to work always, maybe by waiting or something. But now they are saying “buy more”.
As long as autocomplete in Cursor is the best (for now), Github Copilot after the last updates is pretty great. They are working on auto-completions too, and they don’t have these stupid limits. Oh, and they cost you 10 bucks a month. Just saying.
Bu you don’t seem to get it! If slow-requests don’t even go through, you are locking the development for that person. This means that for this to be an even usable product you have to GUARANTEE that slow requests work withing some set time-limit. It is up to YOU to buy compute so that we never end up in this state. You are selling more than you can handle, and that is stealing peoples money!
You don’t feel the least shame when saying that?
Let me translate: “Hey, we have oversold our product! But instead us fixing it, we think it’s better if you pay even more. This way we avoid taking accountability for our mistake, and we still get more money!”
We have now increased our slow requests capacity, so you should not see these errors as frequently anymore.
We don’t believe we have oversold our product, as we do not cap your usage of the slow requests in any way. However, we do choose to reject requests which we know we will not respond to in a reasonable amount of time.
We work hard to balance our capacity, so that we don’t have to increase our pricing but can still offer the same usage allowances in our plan. We could up the price of Pro and increase the capacity of the slow pool further, but that could push out customers who may not afford that higher price, so instead we give you options.
If you run out of fast requests, you can always:
- Pay for more fast requests
- Use our slow requests pool when available
- Use a lower model, like gpt-4o-mini, without any queues or delays
- Add an API key to bill directly to the LLM provider
Despite what anyone, including us, would like, we can’t give away AI requests for free.
We could do what ChatGPT does and force you to use a lower LLM, not even offering you the slow requests pool, and make paying for more fast requests your only option. But we know there are times where you need a better model, but are outside your allowance, which is why we offer the slow pool.
We are not out to make a massive profit on our users, and price very deliberately to allow for every user, from the hobby weekend coder to the career engineer, to get what they need out of Cursor.