Hello! Wanted to give some details on slow pool wait times and why they’ve recently increased a bit…
Anthropic Capacity
We’re working with Anthropic to scale up Sonnet traffic. At the moment we’re in a capacity crunch and Anthropic cannot sustain more TPM (tokens per-minute), which is hurting the slow pool.
Newer Token-Hungry Features
For features like agent, context windows are larger. This means we hit capacity limits faster on the slow pool and the slow pool can sustain less traffic at the same cost.
Immediate Fixes
Wait times are calculated proportional to how many slow requests you’ve used so far, so wait times shouldn’t be bad unless well over your fast request limit.
If you’d like to bypass the wait times entirely, you can enable usage-based pricing. You’ll be charged only for what you use over your 500 fast requests.
You can also enable other models that are not in a capacity crunch. claude-3.5-haiku uses 1/3rd of a fast request. There are lots of custom models for non-agent features, like gemini flash 2.0
We’re also working on a few other, custom alternatives that won’t have the same capacity constraints.
I understand that there is an overload at Anthropic and that they are working on possible solutions. The model you developed was very good, so I imagine it scaled globally and Anthropic’s servers reached the limit of their capacity.
But I would like to know, from your side, the idea is to return to delivering the performance we had before or just placing more tokens paid for use?? Well, if to achieve better performance I need to keep renewing credits, unfortunately I will need to look for new alternatives.
There was no such problem in the early days, slow requests were coming at an acceptable speed. But now this period has gone on for too long and you suggest turning on usage-based pricing. Unfortunately, this is not a permanent solution, we are already paying for the subscription. Slow requests need to be optimized somehow.
I’m really frustrated by this solution. Offering unlimited slow requests shouldn’t mean waiting over 5 minutes per prompt due to individual usage limits—especially when the solution is to buy more credits.
This isn’t the right way to treat paying customers. If service costs are higher than expected, it would make more sense to adjust pricing, introduce new tiers, or update the terms. Slowing down the product and providing poor support only hurts the user experience.
The other issue is that even when waiting 5+ minutes or
More, the amount of times that the agent times out, or comes back with “I’m sorry I can’t do this but these are the next steps” makes it that much more frustrating.
I bought a 1 year pro membership and now it takes me 5 minutes to do a transaction. This is too high. You are still saying to pay. According to your explanation, as far as I understand, using Cursor is illogical. So what you should say is that we are aware of the problem and we will fix it.
If there is a possibility that it will not be fixed, I will look into ways to cancel my pro membership.
I would back this up, I will wait till the end of the week and then file in a request to retrieve my 200$ back. I did purchase an annual subscription in order to save money, first, but also because I trusted this whole project and was willing to finance it regarding the scaling process that should take place. This scaling process is far from being sufficient yet. Last reason was that I wanted to avoid any monthly subscription price increase for the next 12 months.
But the key point here is that without the “Agent” mode, Cursor is no different from competition. All the advice provided here is basically to truncate Cursor from this key feature, either by switching model (and most of us here are using Claude for a reason), or using the “normal chat” window (in which case I can as well use a different tool than Cursor).
Then, second advice is to buy credits but, as most customers, I would not have paid for Cursor if the price was not “capped”. I have a subscription and there are terms & conditions associated to it. If these conditions are not met, I will ask for my money back.
Last but not least, the sentence above “Wait times are calculated proportional to how many slow requests you’ve used so far, so wait times shouldn’t be bad unless well over your fast request limit.” is hard to understand. For me you are basically stating that the more you use Cursor, the slower it gets. If so, the tool is of limited usage, and should thus be of limited price.
Please elaborate on this last statement, I find it hard to get.
Thanks a lot for the hard work and amazing tool you did create, and I hope this is a one-time only event that will not lead me to ask for a refund.
I think business wise it is not a good idea for Cursor because they lose potential customer and therefore money.
But given the context and with this episode in the last couple of days, I also think they lost a non-negligible amount of paying customers.
It is all about where you put your priority, acquiring new customers or retaining loyal ones.
I really think they could have worked around all those aspects with better communication and transparency. It is true that @amanrs stated that Anthropic had capacity problems early, but the first message was not very clear.
This one is better but came too late in my opinion, and also not a good move to ask to pay more as a solution.
Just saying something like “we apologize for the inconvenience, we are working to resolve this as soon as possible, thanks for your patience” with some clear and transparent reasons as to why and some ETA and updates on the evolution, would have sufficed to keep confidence of paying customers and give them time to focus on resolving whatever is happening (Don’t forget a high percentage of your customers are fellow SE and programmers, so we know sh*t happens in our jobs).
That being said, not defending anyone, but if we are being honest, it is a young company that did not exist almost a year ago, young founders, attacking a really hard and highly competitive market at the moment.
In all fairness, we must take into consideration those aspects and give them some slack to work toward making the product better. It is okay to be unsatisfied, and I understand the ones that are leaving, but it is the first big issue that is taking a long time to be resolved, and it is not even 100% their fault.
Also, it is so sudden and big that I suspect some external intervention or attack to harm Anthropic which Cursor is really dependent on, unfortunately.
In the end, I really hope they work around the issue soon, Cursor has still great potential.
just count the number of times they have mentioned buying more credits and compare that to the times they have 1. taken responsibility for not holding up their end of the “unlimited slow requests” part of the subscription. 2. any sort of apology or anything even resemblng that.
they saw a problem at anthropic and instead of freaking out, they stayed calm and came up with a way to expoit this problem ad claw back some of that lost revenue in the form of digital planned obsolescence. Apple is jealous of you
Other question: how much do 500 fast requests approximately cost for Claude sonnet 3.5 for usage-based pricing given you have large context requests of about 500-500-1000 lines of code (~average 5000 input tokens per request)?
For output tokens, let’s estimate a typical response might be around 1000 tokens (though this can vary significantly based on what you’re asking Claude to do with the code):
Output tokens: 500 requests × 1000 tokens = 0.5M tokens
Output cost: 0.5M tokens × ($15/M tokens) = $7.50
Total estimated cost: $7.50 + $7.50 = $15.00
This is a rough approximation assuming:
Consistent 5000 token inputs
Average 1000 token outputs
No additional context or variations
No other fees or minimums
The actual cost could vary based on:
How verbose the responses need to be
Whether you need detailed explanations vs brief answers
The actual token count of your specific code/inputs
Any additional context you include in requests
This shows how Cursor is highly optimal for cost-to-value, I would also add that Cursor intelligently manages the context for prompt iteration, you’ll use a lot more context(and money) using tools like Cline, I would also add that anyone telling another tool does better than Cursor for cost-to-value is highly incorrect and probably a paid actor, do your own research and you’ll come to the same conclusions, are there problems? yes, is Cursor still the best? yes, no doubts
