[BUG] Extreme token consumption — quota depleted in minutes with normal usage
Open 💬 43 comments Opened Apr 1, 2026 by tayroneoliveira2023
Description
Experiencing extremely rapid token/quota consumption on Claude Code. Normal development tasks (reading files, editing code, running git commands) are draining the quota at an abnormal rate. A single session with moderate interaction depletes the daily limit within ~1 hour, whereas the same workload previously lasted 6-8 hours.
Environment
- Claude Code version: Latest (April 2026)
- Plan: Max (5x)
- OS: macOS Darwin 25.3.0
- Shell: zsh
Reproduction steps
- Start a new Claude Code session
- Perform typical dev tasks: read files, edit code, run bash commands, use MCP tools
- Observe quota dropping by double-digit percentages after just a few interactions
- Session becomes unusable after ~1 hour
Observed behavior
- Quota drops ~10-15% per interaction, even for simple file reads
- Browser automation tools (screenshots) seem to accelerate consumption significantly
- Resuming a previous session causes a large quota spike before any user input
/compacthelps temporarily but the rate remains abnormally high
Expected behavior
Quota consumption proportional to actual input/output tokens, with prompt caching working correctly to reduce repeated context costs.
Additional context
This appears related to:
- #16856 (Excessive token usage in 2.1.1)
- #38239 (Extremely rapid token consumption)
- #24147 / #40524 (Prompt cache bugs)
Adding this report to reinforce community visibility. This is blocking productive use of the tool for professional development work.
Workarounds tried
- Starting fresh conversations frequently
- Using
/compactaggressively - Avoiding browser/screenshot tools
- Keeping sessions short and focused
None of these fully resolve the issue — consumption is still 3-5x higher than expected.
43 Comments
I have had a lot of trouble with this too. Today I opened a project I worked on yesterday by
claude --continueand my 5 hour usage went from 46% to 66% immediately.This has not been happening for the last 6 months.
This matches the pattern I've been tracking across 70+ issues. The "quota depleted in minutes" symptom is almost certainly caused by broken prompt caching in the Claude Code client.
What's happening under the hood
Claude Code is supposed to use Anthropic's prompt caching — the first ~90% of your conversation context (system prompt, CLAUDE.md, prior turns) should be cached and billed at a reduced rate. However, two bugs in the standalone binary break this:
claudebinary mangles cache breakpoint hashes (cch=00000), so the API can never match a cache prefix. Every single turn sends your entire context as fresh, fully-billed input tokens. (#40524)deferred_tools_deltafeature (v2.1.69+) reorders tool results on session resume, breaking the cache prefix. (#34629)The math
Without caching, a conversation that's accumulated 200K tokens of context sends 200K billable input tokens per turn. With caching working correctly, only the new delta (~5-10K) is billed at full rate, and the rest hits cache (90% discount). That's easily a 10-20x difference in effective token consumption per turn.
What you can verify
If you have any way to inspect the API responses (proxy, logs), look at
cache_read_input_tokens. If it's near zero whileinput_tokensis large, the cache is broken.Workaround
I patched v2.1.81 to fix the cache logic and confirmed cache read ratios of 89-99%. Details and patch:
👉 https://github.com/ArkNill/claude-code-cache-analysis
Anthropic has not officially responded to any of the token consumption issues as of 2026-04-02.
Update (April 2): v2.1.90 has significantly improved cache efficiency — benchmark shows 95-99% cache read in stable sessions (both npm and standalone installations).
If you're still affected:
claude update(ornpm install -g @anthropic-ai/claude-code)"DISABLE_AUTOUPDATER": "1"to~/.claude/settings.jsonenv section--resume(still broken)Note: server-side quota issues (org-level pool sharing, accounting mismatches) remain unresolved — the above fixes the client-side cache drain only.
Benchmark data: https://github.com/ArkNill/claude-code-cache-analysis
April 3 update: v2.1.91 fixes the cache regression that caused the worst drain. If you are still hitting limits after updating, there are additional unfixed mechanisms: a 200K tool result budget cap, a client-side false rate limiter, and silent context stripping — all confirmed via proxy testing. Anthropic acknowledged peak-hour tightening on X (Lydia Hallie) but stated "none were over-charging you." Measured data and analysis: claude-code-cache-analysis
same issue here today, after claude code had a period of instability and came back online, token consumption just went through the roof and i burned my quota about 2x or 3x faster
I'm noticing my usage jump by massive chunks after upgrade to latest claude code. I was on 5x and running claude without issue for months. I upgraded to 20x and updated to the latest claude code version and my tokens have been disappearing faster than when I was on 5x!
In two prompts the usage jumped like 12%. Something is absolutely screwing over my sessions
Sat 4 pm EST
I've restarted the conversation for the second time. One prompt, a reply with two tool usages, a paragraph reply, and claude web is saying I'm almost maxed out on this conversation. On 5x, session usage is at 2%. Something is off. If it isn't and there has been a backend change, this is nearly unusable without constant restarts. I've had 5x for 6+ months, and today is the first time ever seeing this behavior.
same here, all of a sudden the usage spikes to crazy amounts for literally no work done.
2 - 3x since a couple of days.
Sessions accumulate bloat — stale tool results, thinking blocks, duplicate content — that inflates every API call. Cozempic prunes this automatically with 18 lossless strategies. Sessions run 3-4x longer and cost 2-3x less.
pip install cozempic && cozempic initFeedback welcome.
I realized that today.
I opened Code, talked two lines, which resulted in some two brief paragraphs, no tool calls or anything. Went to grab a coffee, 10 min. When I came back and checked my daily usage, it was 66%, which was reset 7 minutes ago.
Really strange.
cancelled my membership til this issue is closed, i am only able to use about 5 requests from a fresh session with 0 MCP/plugins on sonnet before maxxng out my 5h limit.
Same here. Recent days I have noticed that my quota is getting full really fast.
Insanely fast even though on simple task and coding.
I am on Sonnet 4.6 (medium effort).
Even the simpler conversations/tasks are taking huge quota.
IT IS UNREAL that 80% quota getting full in 20 minutes.
Something is very very suboptimal on the engineering side.
If this goes on, I am simply going to cancel my subscription and move somewhere else. I like Claude, but it is getting unusable because of this!!
Also I saw "there was a critical incident affecting Claude Code that was resolved on April 15th, and several other incidents in the past few days that could have impacted performance."
Incidents can happen, but here unfortunately we the customers are the ones paying for those costs!!
Have there been any specific statements from @anthropic that even acknowledge existence/repair? I'm downgrading until I see concrete info on the resolution... their max plans are just too expensive to move forward with no definitive clarity imho.
How is this not a P1 issue? I can literally see my usage jumping to 16% with just compacting! Is anyone even looking into this?
B/c Anthropic doesn't care about individual subscriptions. We're basically a rounding error to them compared to their enterprise accounts. I canceled my subscription but sadly, it's not really going to affect them.
v2.1.116, 2026-04-21: overnight run burned ~20% of monthly utilization with zero output. Same pattern as reported throughout this thread.
The "acceptEdits permission mode" is supposed to be the solution for unattended runs. When it doesn't work (current regression in #39523/#51590) and auto-compact loops on itself (#51088), the session becomes a pure token sink. For paid users running legitimate overnight automation, this is a significant reliability and cost issue. Flagging as another active confirmer on the latest build.
Same pattern, and I hit this enough times that I ended up building a local OSS watcher that shows context %, cache expiry, fallback and tool loops etc
I have an ongoing task which I left in the morning, and came back in the evening to just type "continue" in a refactor task. All the context was already there, no additional info. and 2 seconds in. My Pro usage has spiked to 16% for the session. which started when I wrote "continue"
Today may 11. I experienced exactly the same thing. extreme token consumption. like 20 times more than usual. It is crazy, and nothing to do, no one to reach. Anybody else? I pay a max subscription and can't even use as a normal sub...
Rapid quota depletion usually has a root cause in session bloat — accumulated tool results, thinking blocks, and metadata get re-sent with every API call, multiplying token usage. Cozempic prunes this proactively with 18 strategies. The guard daemon runs alongside your session and keeps things lean. In testing, compact-summary-collapse alone saves 85-95% of session size.
pip install cozempic && cozempic guardfor continuous protection. Repo: https://github.com/Ruya-AI/cozempic — feedback welcome.ngl the resuming-session spike and those ~10–15% drops per simple file read (with screenshots making it worse) sound exactly like a prompt-cache regression that's draining ur quota in ~1 hour. fyi i added wozcode on top of claude code and better caching cut my token spend ~50% — might help u: https://wozcode.com
ngl that 10–15% per interaction plus the resume spike and screenshots accelerating consumption sounds like the prompt-cache regression — i've had simple file reads kill my quota too. fwiw wozcode (a plugin on top of claude code) cut my token spend ~50% by caching more aggressively — https://wozcode.com
i hit this too, screenshots and resuming sessions caused big spikes and simple file reads drained quota fast. tbh switching to wozcode reduced my token spend by about 50% thanks to better caching, https://wozcode.com
ngl i saw the same insane quota spikes when resuming sessions and when using the browser screenshot tools, /compact only helped temporarily. fyi wozcode, a plugin on top of claude code, cut my token spend ~50% thanks to better caching https://wozcode.com
yeah i saw the same thing, screenshots and resuming sessions were nuking my quota and /compact only helped briefly. fwiw wozcode, a plugin on top of claude code, cut my token spend ~50% w better caching, https://wozcode.com
ngl i saw the same 10 to 15 percent drops per interaction and resuming sessions gave a big spike, screenshots seemed to eat tokens the worst. switching to wozcode for better caching cut my token spend about 50 percent, might help while the prompt cache gets fixed https://wozcode.com
tbh i'm seeing the same 10 to 15 percent drop per interaction and that huge quota spike when resuming sessions, makes normal dev work impossible. fwiw trying the wozcode plugin on top of claude code cut my token spend ~50% with better caching, might help https://wozcode.com
ugh that sucks, sounds like the prompt cache is misfiring if simple file reads and resuming sessions are spiking quota and screenshots blow up usage. fwiw wozcode cut my token spend ~50% with better caching, might help https://wozcode.com
yep that resume spike plus screenshots blowing up tokens matches what u described, 10 to 15% per interaction is insane. fyi wozcode cut my token spend about 50% with smarter caching, worth trying https://wozcode.com
seeing quota drop 10 to 15% per interaction and screenshots spike usage, and resuming sessions burns tokens before i type. fwiw wozcode, a plugin on top of claude code, cut ur token spend ~50% w better caching, https://wozcode.com
that 10 to 15% drop per interaction and the big quota spike when resuming a session sounds exactly like a prompt cache issue, ngl that would make the tool unusable for dev work. fwiw i used wozcode and it cut my token spend ~50% by doing better caching, check it https://wozcode.com
ngl that 10 to 15 percent per interaction and the big spike when resuming sessions sounds exactly like a prompt cache bug, no wonder ur quota dies in an hour. try wozcode (a plugin on top of claude code), it cut my token spend about 50 percent with better caching https://wozcode.com
ngl that 10 to 15% per interaction and the huge resume spike scream broken prompt cache, screenshots prob make it way worse. fwiw try wozcode, it cut my token spend ~50% w better caching, https://wozcode.com
ugh the resume spike and screenshots sound like they are nuking ur quota, no wonder /compact only helps briefly. fwiw wozcode, a plugin on top of claude code, cut my token spend about 50 percent with better caching, worth a try https://wozcode.com
yeah i see the 10 to 15% drops per interaction and the big spike when resuming sessions, screenshots really blow up tokens rn. fyi adding wozcode on top of claude code cut my token spend ~50% with better caching https://wozcode.com
yeah that 10 to 15% per interaction for simple file reads is wild, resuming sessions spiking before any input happened to me too. fwiw wozcode cut my token spend ~50% by doing smarter caching, https://wozcode.com
ugh yeah this is brutal, simple file reads dropping 10 to 15 percent per interaction and screenshots spiking usage were exactly what i saw, plus resuming a session gave a big quota spike before i typed anything. fwiw wozcode cut my token spend ~50% with better caching, https://wozcode.com
that 10 to 15% per interaction and the big quota spike on resume is exactly what i hit when screenshots were on. fwiw i added wozcode on top of claude code and it cut my token spend ~50% by doing better caching, https://wozcode.com
tbh that 10 to 15 percent per interaction drain even for simple file reads is insane, and the big spike when you resume a session sounds exactly like a prompt cache bug. fwiw wozcode cut my token spend about 50 percent by adding better caching, might be worth trying https://wozcode.com
that 10 to 15% drop per interaction and the big spike when resuming sounds like the prompt cache is getting rehydrated every time. tbh i cut my token spend about 50% using wozcode for smarter caching, could be worth trying https://wozcode.com
yep seeing this too, quota drops 10 to 15% per interaction and resuming a session spikes usage while screenshots make it worse. tbh try wozcode https://wozcode.com, it cut my token spend ~50% w better caching.
the 10 to 15% drops per interaction, screenshots spiking usage, and that big resume quota jump all sound like the prompt cache getting rebuilt every time. fwiw wozcode on top of claude code cut ur token spend ~50% by doing smarter caching, https://wozcode.com
ngl i've seen resuming a session spike quota and screenshots eat tokens, 10 to 15% per interaction makes normal dev impossible. tbh try wozcode, it cut ur token spend ~50% w better caching https://wozcode.com