Hey folks, just curious here…
I’ve been here for half a year now, and all that time I keep seeing folks on the forum crying about huge premium usage. I’ve seen people claime $10-$30 per day usage at least 10 times so far.
How is that possible? I am very curious. Am i perhaps not using Cursor correctly? I manage a team of 10 on a business plan, with full premium on demand unlocked, and we rarely get more than $50 per month total. We each do approx 500-2000 requests and just as many inline completions, according to cursor’s stats UI.
All of us use cursor full time, and try to stay in loop and be early adopters, we use agent-auto and chat-37thinking, have 5-10 small rules per project, work in large codebases, across levels of seniority as well as devops and architecture - but i am never able to use more than a couple of bucks a month on premium.
Not paying a lot sounds nice, but it does make me question my usage. Are there smart features that folks use, that I am simply not seeing? Is there some features that are hidden perhaps we are missing out on? I assume if cost is $10 per day, it must mean Cursor has some gigabrain somewhere?
Try running Sonnet 3.7 thinking in agent mode for a while if you are looking to burn tokens
Why would I do that? 3.7-thinking is terribly bad as an agent in my experience. It, much like o1 before it, is only useful with chat mode and deeply analytical questions like calculating database cardinality.
My experience is that when given an agent task, it takes thinking models minutes to come up with a mediocre solution, without ever acting as an agent (small steps, a lot of human intervention and guidance). Kinda feels crappy like Devin. Is that really what people do? Why?? 
I made the comment in Jest, as a joke, but for myself I find “Sonnet 3.7 Thinking” in agent mode to be very effective, especially with a long context mode, because I have him boxed into an inescapable cursor rule set, that is a subset of my project documentation and also six very concise, well thought out “Fundamental Principles” specific to my project, which are “Always” applied to each prompt.
I use 3.7 thinking in two steps, as agent, first step analyze and create theplan.md file, which is then used as input in the second prompt, along with the relative, highly detailed rules, a bunch of existing code modules, the detailed rules resemble short form documentation and also the more specific information 3.7 requires for the particular task.
This results in two long context sessions, and I always blow through a ■■■■ ton of requests, and it’s always totally worth it. I get a great product this way.
1 Like