Cursor burning through added On-Demand monthly usage at an insane rate while returning with more errors

Support Help
1

I’ve been using Cursor for less than a month and it works amazingly, the only thing that is killing me as a solo developer is the cost of each prompt, at first I chalked it up to my prompts not being optimized so I asked other AI for help making better prompts, that yielded better results but the cost of each prompt and how many tokens are being used is insane. I dropped around 60M tokens (diagnosing and fixing a single pretty simple issue because cursor was hallucinating that it had fixed the issue and kept returning to me as such. Eventually plugging everything in my output into ChatGPT, not even the paid model, diagnosed and provided a fix immediately. I have been using ChatGPT to help prompt cursor and that has been working better but the token consumption is insane? Does anybody have any tips or recommendations to make cursor use less token or at the very least more efficiently, I’ve already had to add more spending limit to it around 3 times (twice of which were from the same issue.)

Edit: Decided to check and the single most consuming prompt was around 4M tokens in a matter of 30 seconds to a minute.

4

deepseekv4 is what people use now adays. better models for cheaper.

you can use stuff like opencode, badgr launch,cline and idk any code editor that lets you put in deepseekv4 minimax etc

6

Hey, thanks for the detailed breakdown. The main thing inflating the number in the dashboard isn’t new input, it’s cached context being replayed on every agent step. Cache reads are much cheaper than fresh input, so the raw token count almost always looks scarier than the real impact on your allowance. On https://cursor.com/dashboard/usage each request is split by token type, which is the fastest way to see if a specific session was actually expensive or mostly cached replay.

What really helps reduce usage:

  • Start a fresh chat for each task. This is the biggest lever. If you keep one long thread, every next request rereads all accumulated history, which is why you can see millions of tokens even for a small change.
  • Narrow the request to one outcome and name the exact files you’re working on.
  • Trim rules with Always Apply and AGENTS.md since they get injected into every request. See Rules | Cursor Docs
  • Set up .cursorignore so generated or archived directories never get pulled into context.
  • For routine tasks, you can explicitly pick a cheaper model such as Composer instead of Auto.

One more thing from your description: the agent “fixed it and said it’s done”, then kept running the same context in a loop. Every extra step replays the whole context again. A fresh, narrow chat for the specific issue usually breaks that cycle.

Let me know if usage still looks weird after this and we can dig in deeper.