[Bug] Unexpected high API usage consumption - 100% quota in 1-2 hours
Open 💬 33 comments Opened Apr 29, 2026 by glaforest
💡 Likely answer: A maintainer (wolffiex, collaborator)
responded on this thread — see the highlighted reply below.
Bug Description
Hi, since last week my usage is getting to 100% in about 1-2 hours. I never had any issue and before last week, I was reaching maybe 10% max every session. Im a 20x customer, I dont think this is normal.
Environment Info
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.123
- Feedback ID: 65f3a35c-f384-4367-9299-7f19480436eb
Errors
[]
33 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
I had the same issue today. First time ever with a single prompt hitting the usage limit (I'm a 5x user). I was willing to assume in my own case that there had been some amount of user error on my part (Opus 4.7 xhigh reasoning with a delegagted agent and a resumed task more than a few hours old). But (a) I'm still surprised, (b) we're both here surprised today (so maybe something changed on their end, (c) even if I used all my limit this quickly, it would be nice if there were something in the harness that halted or interrupted prompts where a user consumes more than 50% of their usage in a very short period of time, especially a single prompt or in the first hour. Diagnostically at least, it seems unlikely this would reflect user intent and so flagging it for the user might be supportive.
Ok, I used 4.6 1M since I have a lot of issue with 4.7. I tried to reach the support, but there is no way to talk to a real person.. I hate to say this, but if there is no fix, I will consider switching to Codex. I use 20x because I want to get a large limit.
The issue seems to stem from sub-agents getting stuck but quota/usage increasing and sub-agents doing nothing. no in/out but usage keeps increasing anthropics side. Fix, downgrade claude-code
Looking at the other issues, I was reminded that Anthropic recently did some rate things too (https://x.com/trq212/status/2037254607001559305?lang=en). So, I'm trying to make excuses for Anthropic I guess, but I normally do most of my clauding after 11pm local time and today chose to do something during a peak time. I'd be curious to know if there's something in the rate limiting that's creating surprises for us here
@Mliviu79 Do you know what version I need to downgrade to?
Same issue here... :(
I downgraded to v2.1.121 and it seems to be fine now. Observation sub-agents got stuck but usage kept increasing. nothing else running
@Mliviu79 I have the issue since last week I dont think v2.1.121 will resolve the issue for my part. Lets hope someone from Anthropic react ....
try 2.1.98
check /usage for more insight into your comsumption, maybe check for how much cache you have etc
Just did, will report back when my usage will reset..
I'm running into the same issue. I was running v2.1.117 and started to notice my current session usage sky rocket on a few queries, nothing token heavy. I updated to latest CC and still same issue.
Lets hope its a bug. I love working with Claude Code...
You guys should post your /usage or /feedback hashes if you want ant people to actually be able to look into this
Here is my usage : Status Config Usage Stats
Current session
██████████████████████████████████████████████████ 100% used
Resets 6:10pm (America/Toronto)
Current week (all models)
███████████████████ 38% used
Resets May 2 at 1am (America/Toronto)
Current week (Sonnet only)
0% used
Extra usage
██████████████████████████████████████████████▏ 92% used
$276.87 / $300.00 spent · Resets May 1 (America/Toronto)
I am surfing on my extra usage and it seems to be way better with 2.1.98. Thank you @Mliviu79 . Seems to be a harness issue.. @claude you should look into this.
If anyone can do /feedback and provide us with a conversation id, we'll take a look.
First time I do this : https://github.com/anthropics/claude-code/issues/54799. Hope I did the right thing.
Postimg the usage has nothing to do with incorrect metrics/calculation on anthropic s website. Usage is taken through an api request. Wrong quota usage/calculation on anthropic side will be reflected there too. Presents no useful value/insight. The fact that I downgraded and the issue is gone is clear that there is something wrong with claude-code or anthropic's backend since all it is are api calls. Also noticed that agents get stuck (doing nothing) but usage keeps increasing and I don't have continuous stream is just a text search/text write. So that might be an infinite loop on anthropic backend that adds to our consumption. Just my case!
You are right the problem is the harness (client), not the server. Downgraded to 2.1.98 and it seems to be way better. Thanks again @Mliviu79 .
Downgrading to 2.1.98 worked for me as well. It took me a bit before I noticed, my weekly usage got chewed up with this bug. (weekly reset for everyone? 😁)
I'm just a commented on this post, my own issue https://github.com/anthropics/claude-code/issues/54778 but my feedback id is 9fb6b56c-3a4e-4217-9d72-7c9981f835a2 and 0682ddc7-9bdc-4f26-ab8c-78c2007dd12d
Edit - I see it now—apologies!
@glaforest Thanks for reporting! I'm not seeing your feedback id, it should look something like this
94b7c59a-e350-4c9a-ae7d-3408768e4a8b. Would you be able to repost it? Would love to investigate what went wrong for you.my /usage as a bit more info:
<img width="1124" height="496" alt="Image" src="https://github.com/user-attachments/assets/357a25fa-6f19-4e01-8ba5-62445abca7f6" />
@willbda‚ both feedback IDs captured an empty transcript (filed from a fresh session). To get the actual data:
claude --resume, pick the big Apr-26 conversation, then run/feedbackfrom inside it and paste the new ID here. Also helpful:/usageStats tab and/contextoutput from that session, plus your mcp__xcode server config (the 100X "not found" errors suggest it's flapping).@mhegazy I'm confused because @glaforest 's feedback id is in the original message post. And for my own case I had done /feedback from within the chat. This might suggest another bug or concern. Does /feedback fail when a user is over their usage limits?
Mine just reset, and so a new feedback id: 50138711-fc2d-42ff-93d0-ec96a51d7f4d
@LouisR-P ‚ thanks, your feedback bundle (
fcbe3f8e-) came through with full data. That one looks like 31 normal Sonnet 4.6 turns at ~75K context ‚expensive but not anomalous on its own.The session you described in your follow-up sounds more diagnostic: clean terminal,
/clear, deleted CLAUDE.md, single prompt 40 minutes / 8 lines. Could you/feedbackfrom that session?Also: which model + effort were you on for that one? (/model will show it.) And paste the new Feedback ID here.
following my own thought here -- when my limit reset, I could see that after the feedback call there was usage by haiku. Separate issue clearly, but it might be worth looking into if people cannot submit feedback effectively when the prompt proceeding their feedback uses up the usage limit.
Session
Total cost: $0.0007
Total duration (API): 1s
Total duration (wall): 8m 13s
Total code changes: 0 lines added, 0 lines removed
Usage by model:
claude-haiku-4-5: 290 input, 77 output, 0 cache read, 0 cache write ($0.0007)
The
/feedbackhas a cap on the size. if the cap is met we drop the large parts (including the raw transcript). in this case this is what i needed to know what is going on. so yes there is another bug. thanks for flaging. working on a fix here.No
/feedbackhas nothing gated on limits. it is a gated on size of the uploads to ensure it does not fail in transportyes. thought this is independent. the Haiku call is to get a session name. this does not gate the feedback submission. it happens in parallel and does not block regardless.
I am having the same problem. Hit usage cap for a session inside of one hour using very basic prompting. My weekly usage rates under 30 percent (have used it every day this week), but session usage rates are off the charts. No ability to provide a nuanced problem report and the AI bot Fin is useless. This has happened on April 28th and 29th. I'm a pro max 5M user.
Hit the 5 hours cap after 20 minutes of GSD running an execute-phase that spawned two subagents. The execute-phase had to change 4 files, and it didn't even complete before hitting the limit. Usually an execute-phase took near 8-12% avg of my 5h window in the past days using Sonnet 4.6 with high effort.
The bug also affected the weekly limits: after two days of work, reached 52% of usage, when in past days I was near 25 to 28%: my ideal goal is to not exceed 14% of weekly usage per day.
tbh that jump from 10% to 100% in 1 to 2 hours sounds like a prompt or model change spiking context tokens, check recent commits or env vars. fwiw wozcode cut my token spend about 50% with better caching, might help https://wozcode.com