[BUG] Usage explosion

Preflight Checklist

I have searched existing issues and this hasn't been reported yet
This is a single bug report (please file separate reports for different bugs)
I am using the latest version of Claude Code

What's Wrong?

Major bug, apparently this morning I have used 50% of my weekly allotment, and some limit i have never ran into before called a session limit. i am not working differently than last week when I was able to last all week.

What Should Happen?

Do any usage, see your token usage blastoff

Error Messages/Logs

Steps to Reproduce

do any usage, see massive token allotment reductions on all vectors

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.81 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

No response

Found 3 possible duplicate issues:

[BUG] Hitting much faster usage limits than before #28537
[BUG] Max x20 plan — 5-hour quota consumed in ~10 prompts (started March 14, 2026) #34410
[BUG] Weekly token burn rate increased significantly overnight - no local changes (Max 20x plan) #35030

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

Observing the same behavior on a MacBook pro. Workflow is unchanged, usage is through the roof. Something is wrong.

Likewise on my m3 air

Same here in my linux Devcontainer. I was on v2.0.76 for a while because it worked and I saw no need to update, until recently when the Task Output started screwing up on me and breaking my workflows. Finally did some incremental updating all the way to 2.1.81 and noticed the huge usage spike in the process.

My running theory is that the API now expects Claude Code to expressly send a signal/flag in its request to set the thinking effort level, otherwise it defaults to "high", which then causes the huge usage spike.

Stopgap solution: make sure you've updated Claude Code to at least v2.1.69 (nice), use the /effort command and make it's set to low.
Also, if you're used to Sonnet/Opus 4.5 like I am, you'll need to explicitly specify the full model IDs now, otherwise it'll default to 4.6. 🙄

/model claude-haiku-4-5-20251001    # Haiku 4.5
/model claude-sonnet-4-5-20250929   # Sonnet 4.5
/model claude-opus-4-5-20251101     # Opus 4.5

Hope this helps!

EDIT 2/23/2026 6:24PM ET:

Just found out you can set environment variables in Claude Code settings.json that map the model aliases to specific model IDs so they won't default to latest. Tested in v2.1.81.

{
  "env": {
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-5-20251101",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-5-20250929",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5-20251001"
  }
}

Unfortunately, I run Opus High by default. This behaviour is not due to switching models or thinking, it is not due to any factor on my end at all.

…

On Mon, Mar 23, 2026 at 3:49 PM Chris Bowe ***@***.***> wrote: *cdbowe* left a comment (anthropics/claude-code#37917) <#37917 (comment)> Same here in my linux Devcontainer. I was on v2.0.76 for a while because it worked and I saw no need to update, until recently when the Task Output started screwing up on me and breaking my workflows. Finally did some incremental updating all the way to 2.1.81 and noticed the huge usage spike in the process. My running theory is that the API now expects Claude Code to expressly send a signal/flag in its request to set the thinking effort level, otherwise it defaults to "high", which then causes the huge usage spike. Stopgap solution: make sure you've updated Claude Code to at least v2.1.69 (nice), use the /effort command and make it's set to low. Also, if you're used to Sonnet/Opus 4.5 like I am, you'll need to explicitly specify the full model IDs now, otherwise it'll default to 4.6. 🙄 /model claude-haiku-4-5-20251001 # Haiku 4.5 /model claude-sonnet-4-5-20250929 # Sonnet 4.5 /model claude-opus-4-5-20251101 # Opus 4.5 Hope this helps! — Reply to this email directly, view it on GitHub <#37917?email_source=notifications&email_token=AFYTNHU57NUQYSURCWXKUIT4SGWM3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJRGM4TCNRYGIY2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4113916821>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFYTNHUPQR626PZWBJE4ZSD4SGWM3AVCNFSM6AAAAACW4N6IB2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCMJTHEYTMOBSGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

So what is the point of these issues and the support email if no one ever responds to them or even acknowledges the bug?

I've used 21% of my session limit in five minutes. Something's definitely off with the counting.

I'm on Opus 4.6 in Medium effort mode. Not doing anything crazy. I never hit my limits or anywhere close this quickly.

Definitely an issue. I used 10% of my Max plan with one prompt asking what version of Claude Code was being used.

Same issue, the usage at my company on the Premium plan has been much higher than usual across multiple developers. No models changed at all. Opus 4.6 in Medium effort by default, but even running haiku shows massive quota burn.

I've used Opus 4.6 effort max since it's available on all my prompt (don't judge me) - I'm on a max x5 and rarely reached the limit of 5h. And with the x2 bonus off hours, it's close to impossible for me (in UE).

But now, in less than 2 hours, i reach the max in few prompts. On the same project, same kind of request, same settings, same claude.md ... And it's not like other older disucssion about debating on the business model, here it's a tangible change in calculation of usage.

So how do we get anyone at anthropic to look or care

…

On Tue, Mar 24, 2026 at 7:20 AM Robin ***@***.***> wrote: *Crazz-E* left a comment (anthropics/claude-code#37917) <#37917 (comment)> I've used Opus 4.6 effort max since it's available on all my prompt (don't judge me) - I'm on a max x5 and rarely reached the limit of 5h. And with the x2 bonus off hours, it's close to impossible for me (in UE). But now, in less than 2 hours, i reach the max in few prompts. On the same project, same kind of request, same settings, same claude.md ... And it's not like other older disucssion about debating on the business model, here it's a tangible change in calculation of usage. — Reply to this email directly, view it on GitHub <#37917?email_source=notifications&email_token=AFYTNHTS47NTGKVK4V46W7D4SKDSRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJRHAZDGNJUGI32M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4118235427>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFYTNHV7E5B3J4APMUR66HT4SKDSRAVCNFSM6AAAAACW4N6IB2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCMJYGIZTKNBSG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

Another confirmation: I had a single fresh session with super simple changes - 5h limit reached in an hour of usage, most of it was just waiting on Claude as it was super slow - I only did a few prompts. Extremely abnormal.

@cdbowe

@cdbowe that tracks. Would also explain why it hit everyone at different times depending on when they updated.

I've been watching my usage turn by turn and the drain definitely isn't even. Some turns barely move the needle, then one randomly eats like 8-10%. Without seeing it happen live it just looks like "I did normal work and suddenly hit 100%."

Built a small statusline for this if anyone wants to try: https://github.com/Astro-Han/claude-lens

At least helps narrow down which turns are the expensive ones.

Same issue for me - I have a tracking tool based on tokens - the high bars are from my last weekly session where I didnt max out usage, the small ones post that are this week, and I've just hit max usage - whats worse is the bug also is applying to extra usage (which ive now turned off), but seemingly it cost me £8 for claude to review a github issue. Completely unacceptable - silence from Anthropic on this too. Based on what I can see there is no obvious cause - its almost liek the usage being logged doesnt match the ACTUAL usage.

Apparently this is a feature now: https://www.reddit.com/r/ClaudeCode/comments/1s4idyz/update_on_session_limits/

For me, 01/04/2026, this morning: rate limit 100% exhausted in ~50 minutes. I waited 4 hours and the rate limit was 100% exhausted again with a single prompt.

what is funny is how the "bug" is rolling out at a controlled pace because I notice this several weeks ago, and every day some issue is posted. Since we use a wide range of CC versions, plans or operative systems... from the outside it looked even unrelated, but it is clear that someone is messing around with limits.

what bothers me the most is the absence of answer, and this lead me to theories regarding strategies in place to measure the demand elasticity, in other words: how much we can take before we drop our subscription.

A user in #40584 may have found the root cause for your rate limit issue. Two findings:

The "Rate limit reached" error is generated CLIENT-SIDE by Claude Code, not by the Anthropic API. Evidence: model: "", input_tokens: 0, output_tokens: 0 — no request was ever sent to the API. The client-side rate limiter counts the entire context size against the token budget, so large/long sessions get permanently blocked even when the API is fully available.
The client-side rate limiter MULTIPLIES estimated context tokens by the number of concurrent requests. So 3 subagents × 400K context = 1.2M tokens counted against budget — even though the context is cached and shared. This means as your session grows, fewer concurrent agents can run before hitting the false limit.

See #40584 for the full analysis with log evidence from two independent occurrences.

A lot of the new complains are related to fresh sessions lasting a few minutes, not related with long or large sessions whatsoever.

In my case, yesterday I started a session with PRO plan and I got to the 5-hour limit even before my first compaction and without using agents, just a discussion over a couple documents...

Same issue as all the above, hit an obscene amount of usage modifying some markdown files, are we going to be reimbursed??

I'm also being affected by this suddenly today. I'm on a Max plan and hit my session limit with what would be extremely casual usage for me. As a test,

I opened the Claude desktop app, went to Settings->Usage.
Hit refresh
Type a new, short, and simple prompt
Hit refresh

Each prompt is using somewhere between 5-10% of my session, and that's before Claude starts working. Interestingly, Claude's action usage is closer to my expectations, which is around 1-2% of my session after ~5 minutes of work.

I'm on a MacBook Air, CC 2.1.98 in VSCode 1.115.0

A few weeks ago, when I paid for the yearly pro plan, I could use Claude code continuously for 3 hours at a time before hitting my session limit. Now, a single similar query wipes me out in minutes. I'm now running rtk+cozempic+headless+openwolf+read-once+haiku-for-agents+lower-context-window, and I get about 15 minutes if I'm lucky. Those tools are saving me significantly in tokens to the point where if this were a few weeks ago, I really doubt I'd hit my 5-hour window max. And I'm doing this at 6 pm PST after the peak hours. Contrast this with the FREE tier of OpenCode, which lets me hack for at least 3 hours before I hit limits.

I have been using Codex non stop since sunday afternoon when my claude threshold was reached. claude reset at 5 pm saturday. im working gemini tiers into the mix as well.

…

On Fri, Apr 10, 2026 at 7:19 PM Joey Stanford ***@***.***> wrote: *rinchen* left a comment (anthropics/claude-code#37917) <#37917 (comment)> A few weeks ago, when I paid for the yearly pro plan, I could use Claude code continuously for 3 hours at a time before hitting my session limit. Now, a single similar query wipes me out in minutes. I'm now running rtk+cozempic+headless+openwolf+read-once+haiku-for-agents+lower-context-window, and I get about 15 minutes if I'm lucky. Those tools are saving me significantly in tokens to the point where if this were a few weeks ago, I really doubt I'd hit my 5-hour window max. And I'm doing this at 6 pm PST after the peak hours. Contrast this with the FREE tier of OpenCode, which lets me hack for at least 3 hours before I hit limits. — Reply to this email directly, view it on GitHub <#37917 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFYTNHVP6HGW7C4NGFRWEU34VGMTTAVCNFSM6AAAAACW4N6IB2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMRXGU3TSMZUHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

Same here. tried to contact Anthropic so many time, no answer.

A /context command used in a chat (Desktop app, wrong move as I was in a regular chat) used 41% of my 5h session just so claude answer me "I dont have this command"

I've tried codex yesterday... Their freeplan allowed me to work for 20 prompts, and I have still 31% of my limit this week. I wasn't even able t do that when I was still paying claude in late april

Navigation Menu

Repository navigation

[BUG] Usage explosion #37917

Preflight Checklist

What's Wrong?

What Should Happen?

Error Messages/Logs

Steps to Reproduce

Claude Model

Is this a regression?

Last Working Version

Claude Code Version

Platform

Operating System

Terminal/Shell

Additional Information

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[BUG] Usage explosion #37917

Description

Preflight Checklist

What's Wrong?

What Should Happen?

Error Messages/Logs

Steps to Reproduce

Claude Model

Is this a regression?

Last Working Version

Claude Code Version

Platform

Operating System

Terminal/Shell

Additional Information

Activity

github-actions commented on Mar 23, 2026

tstopak commented on Mar 23, 2026

danielctc commented on Mar 23, 2026

cdbowe commented on Mar 23, 2026

ninjra commented on Mar 23, 2026

ninjra commented on Mar 23, 2026

airowe commented on Mar 23, 2026

newfossils commented on Mar 23, 2026

jreed-Aces commented on Mar 24, 2026

Crazz-E commented on Mar 24, 2026

ninjra commented on Mar 24, 2026

sunaurus commented on Mar 24, 2026

Astro-Han commented on Mar 25, 2026

nubbymong commented on Mar 26, 2026

tstopak commented on Mar 26, 2026

Tito-42 commented on Apr 1, 2026

onomarco commented on Apr 1, 2026

marlvinvu commented on Apr 2, 2026

onomarco commented on Apr 2, 2026

lil-itu commented on Apr 3, 2026

mhetrick commented on Apr 10, 2026

rinchen commented on Apr 11, 2026

ninjra commented on Apr 11, 2026

Norbz commented on May 15, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions