Hey guys, is it normal for claude opus to spend the crazy amount of tokens i’m spending right now?
THis is turning absolutely impossible to maintain…
I barely code with cursor, im not cool with trash code, so i just talk to my Agent in order to decide the best options possible…
I have Skills and Rules for everything (separated, Only 1 rule always applied)
I do not have MCP Servers installed
I cleanup chats, 1 chat for each feature!
I close tabs!
I mention Files and Paths!
I use good and concise syntax!
I have a cursor ignore for every extension, only allowed programming files
I’m using Opus 4.7 since composer just says trash
I’m on Cursor Ultra
The project i work on is on the bigger side, so its one reason… But the context just for “ask mode” is really crazy! And its starting to not worth the cost…
Anyone dealing with the same? I’ve spent hours trying to track usage, context leaks etc, nothing solves this…
Gpt 5.5 has about the same Token Usage
Hey, thanks for the detailed post. We looked into the usage behind this, and the 50M number is real as a gross token total, but it is not 50M fresh input tokens or one raw model call (as you also confirm in the screenshot).
Most of that total is cache reads. That means the same large context is being re-read across a long agent run, and cache reads are priced much lower than normal input tokens. This also explains why you can see similar token totals across Opus and GPT-5.5: the main driver is the size of the conversation/project context being reused, not just the model provider.
Given the size of the project, the most concrete ways to reduce this are to keep runs narrower: mention only the specific files/folders needed, exclude generated/vendor/build artifacts in .cursorignore, start fresh chats for unrelated work, and avoid letting one agent run accumulate a lot of tool output/history.
So this looks like very high usage from long sessions on a large codebase with expensive models, not an error in the token accounting.
There’s a problem on cursor for sure, this project used to be way bigger, i worked with models way more expensive, and this is not normal… This was after the last update only
I have skills, rules, cursorignore, every thing you can think, and im having messages costing me over 50M tokens! The 50M tokens mentioned was a chat with 3 Messages! Barely no Context. How could it have that kind of usage? Why is it making Thousands of requests for a single message? Something not ok with Opus for sure…
There has to be a problem, ive been using cursor for over a year, used models more expensive than opus, and never seen something like this.
Anyways, will be switching back to claude while waiting for a fix… Hope this gets fixed and i get my Tokens back, not nice to feel stolen…
Update:
48M tokens equals about 38M words? 60 books? In 3 Messages?
My project is big, but not absurd big. Been tracking the problem, it seems like Agent Looping is the problem. Cursor is making 500-1000 requests for a single response.
