anthropics
[BUG] Insanely rapid usage limit reached!
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
I know there are other tickets for this (https://github.com/anthropics/claude-code/issues/9094), but this issue is too vital to be ignored!
With 3-4 days of 8 hour usage, I am nearly at my weekly usage limit. I am on the Max plan, which Anthropic has stated should be good for 240-400 hours of usage.... The maths aint mathing.
This is affecting your CORE customers, and should be considered a CRITICAL issue.
What Should Happen?
Claude Code usage limits should be provided as advertised, and communicated. There should be TRANSPARENCY so we can see how our usage is calculated. If there is a formula, provide the formula, after all, your primary demographic are developers.
Error Messages/Logs
Steps to Reproduce
Use Claude Code with a regular Max account and you'll see...
Claude Model
None
Is this a regression?
Yes, this worked in a previous version
Last Working Version
No response
Claude Code Version
2.0.15
Platform
Anthropic API
Operating System
macOS
Terminal/Shell
Terminal.app (macOS)
Additional Information
No response
Found 3 possible duplicate issues:
- https://github.com/anthropics/claude-code/issues/8604
- https://github.com/anthropics/claude-code/issues/9310
- https://github.com/anthropics/claude-code/issues/8742
This issue will be automatically closed as a duplicate in 3 days.
- If your issue is a duplicate, please close it and 👍 the existing issue instead
- To prevent auto-closure, add a comment or 👎 this comment
🤖 Generated with Claude Code
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
If the duplicates keep being closed as a duplicate, we'll just have to keep re-opening them?
I've documented this problem(back when 2.0.13 was released). They have now released 2.0.15 WITHOUT FIXING THE ISSUE.
The problem is, according to Claude itself, the system reminder is sending ENTIRE FILES AS REMINDERS when claude has edited a file. I have confirmed this DOES NOT HAPPEN IN 2.0.10 but DOES happen in 2.0.13,2.0.14 AND 2.0.15. Are you guys going to FIX this or continue to let your customers BLOW THROUGH THEIR TOKEN USAGE LIMITS AT 20 TIMES THE NORMAL RATE?
I just had a strange case, just by opening and closing Claude, it used 1% of current session.
I just had a strange case, just by opening and closing Claude, it used 1% of current session.
Yes because it receives instructions before you start from claude.md and system instructions
I just had a strange case, just by opening and closing Claude, it used 1% of current session.
Yes because it receives instructions before you start from claude.md and system instructions
This is my claude.md file, as you can see it has 48 lines all, not to complicated. What is more strange is that every time I close/open session, in the end after making frontend changes it runs "npm run build", which I wrote in instructions not to do so. It said that I need to specifically say to read claude.md everytime I start a session, because it does not read automatically.
So, it is spending 1% for that. I think it is to much for simple 48 lines of instructions, and it is spending un-neccessary because it does not comply with it later, it again ruins "npm run build" even that this is the first instruction.
The system prompt is loaded every time you start claude code. That + tool description tokens are where your 1%(~20k tokens) is coming from on a simple start and exit. My startup usage is between 18k and 40k depending what Im working on because your CLAUDE.md files are ALSO loaded. Just use /context to see what was loaded before you do ANYTHING when you open Claude Code. It lists it all. ALL of that is ALREADY USED by the time you get to the prompt screen. You can't prevent it from happening. You CAN use your own system prompt now(--system-prompt cli argument) AND you can empty your CLAUDE.md files, but you can NEVER remove the built in tool descriptions so it will NEVER be 0 tokens used even if you just open and close the app.
The system prompt is loaded every time you start claude code. That + tool description tokens are where your 1%(~20k tokens) is coming from on a simple start and exit. My startup usage is between 18k and 40k depending what Im working on because your CLAUDE.md files are ALSO loaded. Just use /context to see what was loaded before you do ANYTHING when you open Claude Code. It lists it all. ALL of that is ALREADY USED by the time you get to the prompt screen. You can't prevent it from happening. You CAN use your own system prompt now(--system-prompt cli argument) AND you can empty your CLAUDE.md files, but you can NEVER remove the built in tool descriptions so it will NEVER be 0 tokens used even if you just open and close the app.
Thank you for the details. But if claude.MD is read, why it does still the things that are stated there not to do? unless I tell explicitly to read claude.MD and not do them
Just because the system prompt is loaded in-memory doesn't mean Claude Code should be making any API calls with that system prompt alone. What would even be the point of that?
You seem to be misunderstanding how TOKEN usage works. The tokens were LOADED on startup of Claude code. THEY WERE USED. PERIOD. DONE. Usage was USED. Even if Claude didnt DO anything with it yet. That was sent to Anthropics servers and COUNTS TOWARD YOUR USAGE. ALWAYS. Open a close Claude Code and DO NOTHING ELSE and eventually you will eat your entire accounts limits. This is just how LLMs WORK.
@acidtech that is not in fact how LLMs work. There is absolutely zero need to send any tokens until the start of actual generation. Even if you take into account prompt caching for performance reasons, there is also zero need for the user themselves to cache the system prompt that is shared by the entire Claude Code userbase. And even if Claude Code does do that for some reason, that's simply bad UX that should have the option to be toggled off.
Why are you SO CERTAIN about something so blatantly FALSE? Your understanding is just hilariously WRONG. PERIOD. WRONG! 🤣
This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
it is still the case. I can barely do 20% of work until I hit usage limit in Pro. I started noticing that mid-November. I use Claude Code across many projects and it is not project specific. I do not have any global settings that would explain that.
