Cursor is INTENTIONALLY Throttling Paying Users

[DT80]

Cursor is knowingly misleading its users. Despite advertising unlimited access, they are intentionally throttling paying customers, reducing their request speeds to a crawl while blaming external factors like Anthropic. But the evidence shows that this isn’t about infrastructure or third-party limitations—this is a deliberate internal policy, designed to quietly restrict access while continuing to take people’s money.

I’ve personally experienced these ridiculous delays firsthand. After not using Cursor for a full week, I opened it again only to find my requests taking 20 minutes to process. And yet, a Cursor employee had the audacity to call me a “super high usage” user. How? How does a week of complete inactivity make me a high-usage user? This isn’t about real-time use; this is Cursor implementing an opaque throttling system and feeding users vague excuses when they start asking questions.

The Anthropic Excuse is a Lie

Cursor’s official response has been to shift blame onto Anthropic, claiming that Claude models are experiencing high demand and that’s why requests are slow. But that excuse falls apart completely when users report the exact same delays with GPT models. If Anthropic is the problem, why are OpenAI models just as slow? The answer is simple: this isn’t an Anthropic problem. This is Cursor throttling its own users.

Let’s not forget, Cursor staff outright admitted in the forums that they are intentionally deprioritizing certain users. A company representative posted this:

“Many users, including yourself, now rely on the slow pool for a lot of their usage, and while this is possible, we have to prioritize those users who do rely on it as a backstop.”

Translation: If you use Cursor too much, we decide when you get access, no matter what you paid for.

This was never disclosed upfront. Nowhere does Cursor say that unlimited access actually means “unlimited, unless we decide you’ve used it too much”. And let’s be clear—this isn’t about a few power users hammering their servers. This is affecting everyday users who are just trying to get work done.

They Knowingly Oversold Their Service

Cursor is pulling the classic SaaS bait-and-switch:

1. Promise unlimited access to attract users.
2. Onboard more subscribers than they can handle.
3. Quietly introduce throttling once capacity is stretched too thin.
4. Gaslight users by calling them “high usage” even when they haven’t used the service for days.

And instead of being honest, they keep moving the goalposts. Their latest response? Telling users to switch to DeepSeek v3, a model that lacks essential features like Composer Agent support. That’s not a real solution. That’s just redirecting the problem somewhere else while avoiding responsibility.

This is False Advertising, Plain and Simple

Users paid for a service that promised unlimited premium AI access. Instead, Cursor has built an opaque system where they secretly throttle paying customers, decide who gets priority access, and shift blame whenever they’re called out.

The fact that they’re still taking money from users while knowingly restricting access is outright dishonest. The fact that their support is ignoring customer complaints just adds insult to injury.

If you’re experiencing slowdowns, know that it’s not a bug. It’s intentional. Cursor is throttling users on purpose, and they don’t think anyone will call them out for it.

They were wrong.

[bifoc34610]

Pricing | Cursor - The AI Code Editor

It’s written on the payment page，If you just need the AI ​​chat function, you should go to the relevant AI official website to pay. The cursor is biased towards code.

• Unlimited completions（copilot）
• 500 fast premium requests per month
• After 500 fast premium requests Unlimited slow premium requests

[DT80]

And what that doesn’t say is that Cursor has intentionally throttled these “slow premium requests” to the point where they are functionally unusable.

This wasn’t always the case—users never had to wait 20 minutes per request before. The slow pool existed, but it was still practical. Then suddenly, without warning, Cursor deliberately slowed it down to a crawl. When users started asking questions, Cursor first blamed Anthropic, claiming Claude models were overloaded. But now we know that’s not true, because GPT models are experiencing the exact same slowdowns. So what’s the real reason?

Cursor oversold its service and, instead of being honest, decided to quietly introduce extreme throttling to control usage. And they’re not even hiding it anymore. A Cursor developer outright admitted in the forums:

“Many users, including yourself, now rely on the slow pool for a lot of their usage, and while this is possible, we have to prioritize those users who do rely on it as a backstop.”

That’s Cursor openly admitting that they are choosing who gets usable response times and who gets thrown into a 20-minute queue. And yet, they still advertise this as unlimited access without explaining what “slow” really means.

So yes, the website says “unlimited slow premium requests.” What it doesn’t say is that Cursor has intentionally crippled them to the point of being worthless. That’s not a service—it’s a bait-and-switch.

Edit 1: bifoc34610 just edited his post and tried justifying this by saying the slow requests are listed on the payment page and that “Cursor is biased towards code.” That doesn’t change the fact that Cursor has deliberately throttled these requests to be nearly unusable.

• This level of extreme slowdowns wasn’t always there—they intentionally made it worse and then blamed Anthropic for weeks.
• Now we know it has nothing to do with Anthropic because GPT models are throttled too.
• If Cursor wasn’t meant for AI chat, then why do they actively promote chat features as part of the service? If they want to limit usage, they should be upfront about it instead of pulling a bait-and-switch.

This isn’t about AI chat vs. coding—it’s about Cursor knowingly crippling a feature people paid for, while still advertising it as unlimited. That’s misleading and dishonest, no matter how they try to justify it.

[decsters01]

I believe that you must use the course a lot, a lot, a lot, a lot more than you can count because honestly my requests have never, at any time, taken more than 40 seconds and look, I use the course a lot, I’ve been using it for about 8 months to 1 year and it has never taken more than 40 seconds. I’ve even created conspiracy theories in my head that my request is not so slow simply because I leave it in privacy mode disabled, but I believe that shouldn’t be it. The question is really, I don’t know what happens, but with me it’s like this and I use it on a large scale that every time I renew my plan, my slow requests end in a maximum of four or five days due to the use of the agent and Sonic now with depsec R1, it’s taking longer to finish, but Sonet is still my main partner.

[Polarising]

Hey,

I understand that you might be feeling frustrated. However, it’s important to recognize that running AI services incurs significant costs for the businesses that host them. I’m not sure how extensively you use Cursor, but I personally use it a lot and have become more flexible with how I utilize different models. I admit that I rarely use agent mode and mostly handle my coding independently, using AI as an enhancer or a speed-up for certain tasks rather than as a replacement for myself or as a second developer— at least most of the time.

Perhaps you could reevaluate your AI model usage and consider whether certain tasks could be handled by more cost-effective options, like 4o-mini or DeepSeek v3. I don’t mean to attack you personally, but if you find yourself running out of fast tokens quickly, it might be worth reassessing your usage of AI or acknowledging that $20 a month may not be sufficient for your needs. If you feel that Cursor is being unfair, you might want to try using your own API keys and compare the costs to the $20 you pay to Cursor.

That said, I can imagine that the throttling you’re experiencing likely indicates that your usage is quite high, which may be costly for Cursor. You might want to check out one of Theo’s recent video (https://www.youtube.com/watch?v=PmVONwkYsg8), where he discusses the financial aspects of running t3chat with a 8$ monthly subscription and how certain users made them lose money as they cost multiple hundred euros compared to the average user .

The fact that Cursor is only $20 while offering a wide range of features, many updated and seamless AI integration is quite impressive.

Lastly, as already pointed out, whilst they haven’t specified how fast “slow requests” are, they have explained how their system works. If your usage is extremely high, it’s only fair to consider that other users who are also paying $20 shouldn’t have to wait while your requests consume a disproportionate amount of resources. The slow-requests basically put you in a queue with other users, where I assume some way of strict-priority queuing is used, where users with the lowest usage get prioritized first, and users with high usage last, providing a fair share for everyone, and quicker response times during low traffic times.

Again, I hope this perspective helps you find a solution that works for you.

[thejamespower]

Very nicely worded and this is a serious issue.

I now want to get a refund, because all of a sudden in the last week Cursor has become unusable. I pay for Pro and additional premium requests and it’s completely unusable now because of the apparent throttling. After my first couple requests, I get continuous timeouts

[danperks]

Hi, thanks for your post on this.

To clarify, you will have to queue longer for the slow pool the more you use it. The phrase “queue” may not be as relevant anymore (used to be a queue), but now it is an artificial limit on how quickly you can execute slow requests, which as I say, gets longer the more you use it.

We do have a cap, and while I don’t know the exact figure, it is much lower than 20 minutes, so there should never be any time you have to wait 20 minutes for one request.

Regarding the queue for the models, this is a bit more complex. Due to the ongoing issues with Anthropic models, these requests will slow down much quicker than the GPT models. While my numbers are not real and are just to demonstrate the point, an example could be that it would take 500 Claude requests to reach the maximum wait time, but could take 1000 GPT requests to do the same.

However, the calculation here is shared, so while you would have lower increments of wait times with the OpenAI models if you have reached the top limit for the queue time, you will have the same queue time for both models. This is so you can’t exploit the system, and get 1500 slow requests with lower queues.

Regarding your complaints about a lack of transparency, we don’t publish the raw logic here to avoid people trying to exploit it, but we make no effort to hide or obscure the system in place here.

While I won’t delve into specifics (unless you would like me to in this thread), but you are easily in the 99th percentile of usage of the slow pool, hence you will see the highest queue times. Most users never have queue times like the ones you are facing.

If you are unhappy with this, feel free to email us at hi@cursor.com, and we will be happy to cancel and refund your subscription.