Thanks for pointing out where to look- this is insane:
9M cache token reads for 148k token input? That’s 60x.
I’m afraid the content mentioned in this link can answer your question. The developers of Cursor ensure the accuracy of the model by consuming tokens in a crazy way, even when it’s sometimes completely unnecessary.the system prompt word “maximize context understanding”
IM getting this issue too. I just used 4 million tokens when asking it to fix some tests. My whole codebase isnt even 4 million tokens.
Totally free and totally useless and totally trash.
I would remove this mode altogether, I think it is used only because of the desperation of the new “pricing system”. Please return the old dashboard with the remaining 500 fast queries and the ability to continue using paid models in the slow queue when the limit is reached, it was very cool
auto mode very often uses GPT 4.1, it is an absolutely useless model for coding, if I had AT LEAST been given the opportunity to choose claude 3.5, I would have been less disappointed.
2 Likes
I am building the R-Code AI IDE i just need 50 sponsor because i am solo developer and it has been 15 days i am customizing the vs code fork to make it first AI IDE that will have full context for the model no useless token usage no useless files sending in context LLM will have always your project context even after every change it never breaks your code i will launch it soon so stay updated i am 100% sure you guys will love it because i am using my own AI IDE and believe me its really amazing i am so excited yo share it with you all i am collecting funds for hosting and finding the solution for payment because stripe is not available in my country
1 Like
Alas, it was only due on the 26th, but on the 19th it suddenly showed that it was used up. I tried the AUTO mode, but it was terrible. Now I don’t know what to do when I open the cursor
1 Like
The current pricing method is simply a rip-off. I paid $60, but I can’t even use it for 5 days! And I can’t see my usage clearly in the background. Have you changed the product manager? If not, I suggest firing him.
1 Like
Yup, they were either editing comments where you are pointing someone to their competitors, or just deleting it. And no, it wasn’t buggy AI doing it, it was person with name as surname as you get the notification when it happens.
2 Likes
I apparently cannot make my own thread evidently, but here’s a screenshot of me paying $300 for 2 million token output because of the cache.
EDIT:
I just noticed, there is 0 cache write yet there’s almost 44 million cache read. What cache is even being read if it’s not written prior?
2 Likes
I will throw in here a very simple example of how bad the situation is:
I just asked Claude 3.5 to write me a gitignore (34 lines) for a small website directory
70 cents approx. Insane. Senior devs earn less than that.
2 Likes
I am sorry, but you are making the system much, much worse day by day. I have been paying cursor for a very long time and day by day I have seen the system getting worse and worse. It is still getting worse. Before that, it was unlimited but with hourly limits. That was really good and I was agreeing, but now it is restricted to direct token and month based. So when it’s over, it’s over and a whole month is wasted. That’s not right, you’re going to lose a lot of users and I’m not going to pay anymore. You’re doing it wrong.
I just worked one hour to fix a small issue, it costed me 30€, this is just insane.. no max mode, only regular thinking model
The longer this goes on, the more people, including me, will be warning people against using Cursor. Every person I tell, they will go on to tell more. The Cursor name will be dead to many people, it will be near impossible to shake the untrustworthy status.
Gaslighting and glazing from the official spokespersons, or just complete silence. That’s not a company I want to help succeed.
9 Likes
Hey, for your info, people are not just unhappy they are literally ■■■■■■-off. I think the death of cursor is very near and you know it too, that’s why you guys are squeezing in the last dollars. You are pointing them with some made up numbers(I believe), how can you say these are not just made up numbers? So you were offering 1000 or 2000 dollars per month of model usage from the beginning for just 20 dollars and now you suddenly decided to charge people the actual 1000 or 2000 dollars. It doesn’t make any sense except for cursor’s death angle.
3 Likes
This is good, I didn’t know this before. I am buying this today. No more cursor.
I asked CC today to setup Capacitor for my mobile app. It did it quite accurately. I managed to stay without tokens for one full hour. Then I asked it to set up native Android app (just PoC, just to log in using my own backend) - and it spent maybe 20 minutes setting up all the dependencies but did it. Then I asked it to use ProJet, and it did it accurately. Now I have tokens but am just tired. For 17$ per month when subscribing yearly.
Wow, dang, that is good to know. I’d never heard of this limitation but after researching apparently it’s a well known thing for all LLMs. This explains (for me anyway) why a simple request can consume massive tokens, if it was made on the tail end of a long chat.
So best practices seem to be:
- Use auto mode when possible
- When using big boy models, keep chats scoped and short
It seems like a lot of the outrage over the new pricing is happening because the user base was accustomed to “1 request = 1 request”, which applied universally even if that one request was to a premium model and part of a very long-running chat with massive context. I can also see how Cursor was probably losing a lot of money in that setup 
Still, as others have mentioned, publicizing some sort of “best practices to conserve token use” guide would go a long way to rebuild user trust.
1 Like
Hey everyone!
So the team have looked into this, and we think we’ve figure out whats going on (even I didn’t know this until today)! To be clear, we don’t currently see any bug or issue with how token usage is being recorded within Cursor.
The way that LLMs work (in all apps, not just Cursor!) is that every time a tool call occurs, the whole conversation prior to that tool call is re-evaluated by the LLM. As that section of conversation has already been seen by that model, these tokens count as ‘Cache Read’ tokens, which are the cheapest type of token.
This does mean that, in long threads with lots of tool calls, the cache read tokens can be much higher just one run through of the conversation! As an example, in a response with 10x tool calls, the first chunk of conversation would have been evaluated 9x times as cache read tokens. This would be the case in any AI application that follows this tool-call pattern and is unfortunately just how LLMs handle tool calls right now.
We’re hoping to bring some better in-app features for seeing your token usage and managing it soon, to help make the most of your monthly usage!
9 Likes
You don’t see any bugs? Maybe it’s not a runtime bug, it’s a context management bug within Cursor. I’ve already ran up $80 of costs just from using Cursor for like an hour or two. Please roll back your build or payment model to the previous version and spend a little bit more time evaluating what just went wrong because this is not realistic in any degree.
Just create a brand new Cursor account not tied to your organization for testing purposes, and prompt it for 1-2 hours using Sonnet 4. You will see what everyone is talking about.
2 Likes
Yes, they were subsidizing $10,000 of token usage per user per month until they just decided to stop very recently. Great business model right?
1 Like