1 Like
I’m so upset about this. Every single small instruction burns 100k token. Then the Cursor team comes and says “ah, it’s Cache Read, 10x cheaper.”
OK THEN, 10K TOKENS FOR A HELLO?? wtf guys? What have you done? Doesn’t Cursor unable to grep files snippets anymore? Is it really necessary sending a huge chunk of the codebase just to solve a simple instruction?
I disabled all MCP, tools. No skills files. Just a simple 30 lines AGENTS.md. And still 100K tokens for a single message chat with simple instruction.
What a shame. Pro+ plan going to waste in a matter of 5 days
2 Likes
The same problem - the Pro+ plan ran out in just two days, even though it used to last me more than a month.
take a note that context window, even single Hello can include
- Rule that always active or not
- System prompt
- Tool Call
- Input, Output
Here example i use MCP (98k) and without MCP (22k) there are big different in here. and as you can see its less than $0.02
to make it more justified, i used Codex in Cursor, with just hello. Codex eat up to 20k context windows. almost same with Cursor without MCP right?
And here other example with Claude Code in image.
Hope you get it
Learn about context window:
Hey, I don’t think you read this paragraph
I disabled all MCP, tools. No skills files. Just a simple 30 lines AGENTS.md. And still 100K tokens for a single message chat with simple instruction.
Still 100k cache read for a single task (not a hello, but simple task)
2 Likes
But the most important thing. And the reason why so many people are complaining:
- Months ago a Pro Plan ($20) lasted about two weeks. Now the Pro+ ($60) lasts less than a week
Something change for sure
1 Like
as you can see in previously post, if you see image claude code CLI, Their System Tools is around 11K, System prompt 3k and memory file is 11k, than message only 8 Token. same with “Hello”. that why its eat almost 20k.
for case 100k token, to be able get better debugging and if you believe its bug. you can report on Bug Reports - Cursor - Community Forum with
-
Request ID
-
Cursor Version
without more data, screenshot or other information. its hard to guest what happen.
Hope it can help you
Also worth noting: those “super-smart” models—like GPT-5.3 Extra High or anything on par with Opus 4.6 and beyond—tend to burn way more tokens. They think longer and deeper, which eats context fast and eat up usage.
Example hello with Codex xHight without MCP, and as you can see its 50k. Here benchmark for token usage:
the 100k+ tokens for a simple message is mostly the system prompt and context that cursor sends before your actual message. that includes:
- cursor’s internal system prompt (instructions for the AI)
- any
alwaysApplyrules from.cursor/rules/ - user rules from settings
- AGENTS.md if present
- codebase indexing context (cursor pre-loads relevant file snippets)
- tool definitions (even if you disabled MCP, cursor still has its built-in tools)
the “cache read” part is real though. most of that 100k is cached between messages in the same conversation, so subsequent messages reuse it at a fraction of the cost. the first message is always the expensive one.
@dbdbdb re: your point about disabling everything and still seeing 100k, that’s because cursor’s own system prompt and tool definitions account for a big chunk. those can’t be turned off. the codebase indexing also adds context even without explicit rules.
one thing you can try: open a completely empty folder with no files, no rules, no AGENTS.md, and say “hello.” if it’s still 100k+ then you know it’s purely cursor overhead. if it drops significantly, your project’s file context is the culprit.
Empty new folder, Ask mode Gemini.
Hello - 18k tokens
Hi (even deleted 1 short user rule just in case) - 13k input
2 Likes
you say “hello” then cursor says
“the user said hello, what could this mean? is he saying hello to the sky? does he question his own existence? is the user really a bully and picking on me? why would he say hi to me? the user requests a response i think, let me consider a response. Well all they said was hello the user gave no context. maybe hello means goodbye, no hello would mean purple not goodbye. The user wants me to completely delete his database so he can live in hell…o… no the user wants precious… my precious….
cursore “hello user, ive gone ahead and deleted your monorepo and made what remnants are left public on hugginface while ordering you an apple mini”
yes, this is common. You can read my answer and @nedcodes answer that explain why this happen. This majority Agents Tool have system prompt. Example Claude Code:
1 Like