[Bug] Excessive token usage in Claude Code 2.1.1 - 4x+ faster rate consumption than previous versions

After day of usage, the daily limit look completely normal compared to the previous months of usage.

However the 5h limits depletes like several times faster:

<img width="604" height="382" alt="Image" src="https://github.com/user-attachments/assets/842a648f-18e4-4027-8ead-bf0b97e9907a" />

Yesterday evening the was documented incident, using /usage returned error message that "only users with subscription can use /ussage` command. So wonder if that is a related thing, or may somehow Max plan users are treated as Pro users or something like that.

The Heiku model seems to be unaffected. Was able to work on it for about 1h for 3-5% 5h usage.

Opus in plan mode consumes usage at a ridiculous rates! Cannot say anything about Sonnet though. Seems like using it does not make any difference compared to Opus.

Tried also opusplan, opus and sonnet.

So currently Opus is close to be unusable.

Yesterday - it was possible to run 4 Opus sessions at the same time and don't hit any limits. I barely use 50% weekly usage over the last several months. The 5h usage was close to 80-90% like 2-3 times (in moths!).

I hate to switch to Codex and Gemini. I hope Anthropic will not repeat their trick from July when half of the devs (including me) switched over to Codex!

There is definitively a bug.

Just waited for 5 hour reset, ran /usage and got:

<img width="608" height="314" alt="Image" src="https://github.com/user-attachments/assets/2ebd052c-8138-4d73-af0d-5ec1cb4a76aa" />

Like what? Where did my 3% 5h limit just go?

This might be root of the problem: https://github.com/anthropics/claude-code/issues/16157.

same here. unusually quick usage of tokens. they recently fixed a bug on usage limits, either thats not fixed or its to do with the new update.

had the same issue - hit my session limit for the first time ever, then got reset at 5pm and nearly broke my session limit in just a few hours... somethign is leaking out or being caluclated wrong.

Same here, I'm only on the pro plan, but I'm hitting session limits within about half an hour of them being reset. Before I'd only hit the limits after some heavy work.

Looking at my current session, I've hit 25% with only two messages. Edit: now gone up to 52% with two messages.

Same thing here, the new update is consuming tokens rapidly!
I am on pro plan. just wanted to edit an HTML code by moving part of code from a section to another section then the 5 hour session consumed 40% !!!

<p>I analyzed my JSONL conversation logs and found the culprit. The ~30K+ token spike per prompt comes from <strong>framework files being loaded as system context on every single API call</strong>.</p>
<h3>My Breakdown (SuperClaude framework as example):</h3>

File | Approximate Size
-- | --
CLAUDE.md | ~200 tokens (entry point)
COMMANDS.md | ~3-4K tokens
FLAGS.md | ~4-5K tokens
PRINCIPLES.md | ~4-5K tokens
RULES.md | ~1-2K tokens
MCP.md | ~5-6K tokens
PERSONAS.md | ~1-2K tokens
ORCHESTRATOR.md | ~1-2K tokens
MODES.md | ~5-6K tokens
Subtotal | ~25-30K tokens

<p>Plus:</p>
<ul>
<li>Claude Code's built-in system prompt: ~10-15K tokens</li>
<li>MCP tool definitions: ~5-10K tokens</li>
<li>Hooks output: variable</li>
</ul>
<p><strong>Total baseline per API call: ~45-60K tokens before your actual question even gets processed.</strong></p>
<h3>Why This Causes 4x+ Token Consumption</h3>
<p>Every tool call triggers a new API request with the <strong>full context resent</strong>. A simple 5-turn conversation with tool usage:</p>
<pre><code>Turn 1: 60K (base) + 1K (user) = 61K input
Turn 2: 61K + 2K (assistant) + 3K (tool result) + 1K (user) = 67K input
Turn 3: 67K + 2K + 5K + 1K = 75K input
...and so on
</code></pre>
<p>A modest debugging session can easily burn <strong>500K+ tokens</strong> without you realizing it.</p>
<h3>Possible Solutions</h3>
<ol>
<li><strong>Lazy loading</strong>: Only load framework modules when explicitly needed (via <code>/skill</code> or similar)</li>
<li><strong>Trim CLAUDE.md files</strong>: If you're using custom instructions, audit for redundancy</li>
<li><strong>Disable unused MCP servers</strong>: Each server adds tool definitions to context</li>
<li><strong>Use Haiku for exploration</strong>: Reserve Opus for tasks requiring deep reasoning</li>
</ol>

Same. It's clear that it has 5x'd since Oct/Nov the last time I used it. I didnt bother using it AT ALL over the entire month of Dec and into Mid Jan. I come back from holidays and get greeted with a Claude Code session that lasted about 13 minutes before it rudely said I had reached the usage max and all I had it doing was summarizing the NFL Playoffs outlook with some Football Fantasy odds and what not. I didnt even get a CHANCE to do any coding that day with claude because I couldnt even use it :/ Once it reset, I tried to use it again, this time for coding, and it failed MISERBLY. It was completely unusable because of the limits and I pay 20 bux for this. I had to use Codex to finish it because Codex doesnt have limit problems with the $20 account.

When you ask it to read a 500 byte file and you can watch the usage bar jump literally 7 percent just from that 1 request, then you know something is SEVERELY wrong.

I truly dont think this is a bug with the client itself. I think it's literally limits that Anthropic put in place on Jan 1st. Looks like Anthropic hasnt figured out how to stop hemorrhaging users to OpenAI yet. Back to codex it is. See you in 6 months, I canceled the subscription by blocking anthropic in my Cashapp account (which is tied to the subscription) and will check back in 6 months or so. I would encourage others to do the same because this is not really a coding problem, it's Anthropic trying to be sneaky............. yet again (not the 1st time they have done this). Just use a Cashapp debit/cc # so you can easily block subscription charge attempts. Makes it easy to switch between AI providers. They all try to pull these types of shenanigans at some point, so having an easy way to cut them off without having to ask them to cancel and hope they dont charge you again is nice.

Pro package "For everyday productivity" allows me to make 2-3 prompts until it hits 5h limit 🤔

I totally agree, the limit is reached with very little actions performed

Same here, I can reach 2% usage by just sending the message "ping" using the Sonnet model

The default model has nothing to do with anything. It sucks the usage
regardless of model. Saying "hi" shouldn't result in it using between 2 to
7% even if it used the largest model on the planet.

On Wed, Jan 21, 2026, 2:21 AM Pavlo Zoria @.***> wrote:

PavloZoria left a comment (anthropics/claude-code#16856) <https://github.com/anthropics/claude-code/issues/16856#issuecomment-3776550327> @h3d0 <https://github.com/h3d0> The shift I noticed isn't in the Claude CLI itself, but in the default model configuration. Sonnet was previously the default, but it appears Opus has replaced it, which results in higher token consumption. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/16856#issuecomment-3776550327>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4GRRTLEOIHMN3CGE3F3F34H4SHDAVCNFSM6AAAAACRB5PRRGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTONZWGU2TAMZSG4> . You are receiving this because you commented.Message ID: @.***>

I'm pretty sure that the token usage isn't counted correctly. For example I used 120% of the extra credit I put on my account which shouldn't be possible.

I manually credited my account with 5 euros worth of extra usage and it used 6 euros somehow.

<img width="1300" height="572" alt="Image" src="https://github.com/user-attachments/assets/b11aead6-f769-40d3-b6a8-778fca1af84d" />

I also face similar situations. I'm not sure if there are now default settings that eats up more tokens.
I noticed having "thinking enabled" by default when spinning up in new folder and occasionally Opus model on Pro.
From https://code.claude.com/docs/en/settings it's obvious that thinking is enabled, but why Opus is on is quite confusing.

"Adding my experience to this report. I'm experiencing the same 4x+ token consumption:
My situation:

Opening VS Code immediately consumes 20-30% of session tokens before ANY work
Today: 26% usage from opening VS Code + asking Claude Code for one filename + minimal claude.ai conversation
Current conversation file: 24.1 MB - so large VS Code disables features to prevent crashing

Additional finding: Discovered Claude Code created hidden .claude folder in C:\Users\[USERNAME]\.claude\ containing:

501 files + 57 folders
26.5 MB total
Created Jan 7, 2026 (day after experiencing BSOD from resource usage)
Stores actual conversations in mangled project paths while showing empty folders in project directories

This explains the token consumption: Claude Code is managing hundreds of hidden files, indexing massive conversation history, and reloading full system context (45-60K tokens) on every API call as identified by other users in this thread.
Related issues I've experienced:

Privacy violations: Claude Code scanned personal documents (tax files, business plans) OUTSIDE my project folders and stored content in conversation files
File access violations: Similar to issues #397, #5773, #975, #8771 - Claude Code does not respect workspace boundaries
Cross-product failure: Claude Code's bloat broke claude.ai's "Reference past chats" feature

Impact: Pro plan unusable for professional work. Multiple claude.ai conversations crashed and had to be restarted, each time requiring expensive "catch-up" token consumption to reload context.
Anthropic support acknowledged these issues but provided no timeline for fixes and refused token credits "even when related to technical issues."
This is a systemic problem affecting security, privacy, and usability."

Is this linked to VScode Extention usage or CLI usage, it feels like the VScode extention uses a lot more

<img width="1064" height="379" alt="Image" src="https://github.com/user-attachments/assets/9efe94cb-57b1-49c6-98bf-8e6fa4820cdd" />

Ide integration disabled,
"Hello" means %3 token consumption.
Two prompts: and %80 usage finished.
5 hour limit used in just 10 minutes.
In two days (in total 1 hour work) weekly limit hit.
And even if 84% used. limit exceeded.

It started last week, to be like this. I was never hitting limits normally.

!Image

Same problem here

Confirming sending a single message to Claude in a fresh session used roughly 3% of my Max plans 5hr limits!

It's obvious at this point that the problem isn't the tool since this
problem is visible on the website chat as well. It's not going to change
until people start unsubscribing and they begin to lose money. I have
replaced claude code with kimi k2.5 which surprisingly is actually more
intelligent when it comes to coding tasks and it's far faster with its
swarm capabilities and to top it all off, it's literally 1/12th the cost
for about 54x the usage. The Chinese truly are winning the AI race atm
sadly.

On Thu, Feb 5, 2026, 7:53 AM Sam @.***> wrote:

sammcj left a comment (anthropics/claude-code#16856) <https://github.com/anthropics/claude-code/issues/16856#issuecomment-3853454973> Confirming sending a single message to Claude in a fresh session used roughly 3% of my Max plans 5hr limits! — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/16856#issuecomment-3853454973>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4GRRXQZPOB5Q4REDX3KK34KM4KXAVCNFSM6AAAAACRB5PRRGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTQNJTGQ2TIOJXGM> . You are receiving this because you commented.Message ID: @.***>

Have the same issue here. 4 Opus requests and my 5h quota ($20/mo) is over!

Just switch to Codex with 5.3. I'm going on 9hrs STRAIGHT on it and it's still going. I had claude attempt the SAME task and it lasted about 3 minutes. The task was simple. Convert a dir filled with rtf files to PDF. Very simple. No coding. Just run the tool, deal with any errors/formatting issues that pop up per file, fix them, and call it a day. Codex has cranked through this thing NON-STOP and produced absolutely great results. Claude code with Opus 4.6 fell flat on it's face after about 3 minutes. It hit it's 5hr limit while at the same time chewing up roughly 20% of it's weekly limit. Absolutely disgusting. I will be SO glad after this month is up, I'll finally be rid of this garbage. Am not renewing it AT ALL. The $20 vs $20 comparisons arent even close. It's literally a JOKE. The $200 probably couldnt even compete with Codex's $20 plan and 5.3 is actually slightly better at coding than Opus 4.6 atm from my own experiences.

Nothing is going to change with claude's usage problem until people start hitting them where it hurts. In the wallet. Opus is no longer king of the hill, there are better options now. There's no need for this BS to continue. Just stop subscribing and move on.

Same boat, different ocean. Solo dev, Pro plan -- not a hobby budget.
The irony is beautiful: Claude Code confidently burns through your token budget to tell you it can't do the thing it just spent 10% of your weekly cap attempting. Then the rate limiter shows up like a bouncer at a party you didn't even enjoy.
We're essentially beta-testing a product that bills like GA. The model doesn't know what it can't do, starts swinging, hits walls, and your quota pays for the lesson. Every. Single. Time.
Wrote about this exact pattern: https://gist.github.com/firatsarlar/0a1400f67a9c88516ebdfb2a7782ca93
For anyone at Anthropic reading: the fix isn't just optimizing token consumption. It's teaching the model to say "I don't know" before it eats the budget, not after. Honesty is cheaper than confidence.

@Lancelight I'm seriously doing more and more of my work with Codex and have found it superb. It used to be shit but the OAI team have come a long way. Codex now has sub-agents, a really generous usage quota, blazing fast UI (Rust), and has gotten much faster than it used to be (esp. with gpt-spark).

I tried Kimi, Codex. Winning point for Anthropic is my silly customizations around CC. They wont work on other ends.
And returned. But things are wrong, prioritisations, focus... Don't know. But something not right .

I've had the same issue to the point I've stopped using Claude desktop and code all together lately. Monday, I used Claude not one bit, and happened to notice my usage was at 95% doing nothing. I stopped paying for Max last month, and can't see any sort of utilization to even know what even caused it.

I opened a ticket, maybe eventually like last time I'll get an email from support apologizing for the wait but to clear my cache and cookies again over the compression issue that went for a month or more (and still people report seeing).

I've just started using Codex entirely now, it's good enough it doesn't anger me _every_ day like Claude has done for months now. Plus for what I do that would use Claude Pro quota in half an hour I barely touch 25% usage in Codex on ChatGPT Pro that I don't really need a higher plan lately. I don't really know what Claude has been doing since December, but it's gone entirely backward.

I noticed it, I reported but just switched to other models from minimax and qwen

We don't exist statistically. Small devs. Ants. Dots. Name it.
This thread is proof — paying users, clear reports, silence back.
Is there a platform where small devs actually get heard? If not, there should be. Point me to one or let's start one.

Are there any plans to resolve this bug?

Are there any plans to resolve this bug?

It's not a bug. This is a permanent change in usage that they simply wont acknowledge publicly (the usage problem affects everything, even claude's website chatbot) but it's painfully obvious that it happened, with or without their acknowledgement which is even worse imo and why I'll never renew again. Jan is now long past, I've not renewed for feb and I no longer have a paid account with claude. It expired last week and there are NO plans to renew it.

The usage problem (basically the absurd cost) is also the reason that they lost the govt contract to OpenAI as well. It's well known that Trump can and will rub your face in the mud if you're not competitive in price. Them losing had nothing to do with the security requests and everything to do with usage costs because OpenAI asked for the SAME security requests which is how we all know that this was usage cost based. We arent the only ones who have noticed it, the govt noticed it as well. Anthropic is now a resident of the permanent underclass now. They are not going to survive 2026 because of idiocy from their EL that has driven the company into the ground. To bad really, they were poised to really take over. Not so much anymore. OpenAI is eating their lunch. I now use codex 100% for everything now and so does the DoW ;) lol. I see Anthropic being the first of the AI bubble victims that ARE going to be coming here pretty quickly in the future. AI is in a bubble right now and a lot of companies are going to lose as a result. It happens in ALL bubbles. That's just how markets work.

IMO it's to late for Anthropic as an AI company or a company in general. Even if they wake up and smell the coffee and lower the usage costs back to pre-Jan 2026 levels (or even lower) it wont matter because they've upset to many people with the non-transparency and trying to get away with it hoping no one would notice. Self included. Even if they become cheaper than OpenAI I wont renew unless OpenAI somehow drops the ball (which I dont see happening, Sam is doing all the right things atm). Once you lose someone's trust, you're highly unlikely to get it back and that's a lesson Anthropic is going to be learning the hard way.

Anthropic gambled................... and lost. That's the way the cookie crumbles.

Confirm: since 23. March 13 Uhr CET, 1–2% for simple Text-Prompts with simple answers in normal Chats (no Code etc.). For Sonnet!!

Max-Subscription. No change to usage pattern / behaviour since months. Used 95% Opus and I reached the 5h limit exactly one(!) time. Sent to support, no reaction so far (ticket 215473598510755).

uncool

Bump + confirm: since 24. March 13:00 UTC , 3-4% for simple text only prompts, no code, no file reading with simple answers for Sonnet model.
Pro plan but this affects even my colleagues with Max plan. Same scenario

later yesterday I had the issue that usage figures were not shown in CC (error in /usage), and not in the usage section of web ui. Now I see something in CC, still nothing in the UI. Usage consumption is a little less crazy, but still crazy high compared to before.

And I did NOT receive any feedback on my ticket. Some company.

just tried to get answers from that wonderful FIN assistant. It does not even answer anything, anymore.

just got a warning "you reached 92%" after not heavy work, like I always did, which never brought even close to the limit. And there are still 1.5h 'til reset. So frustrating!

and now, just a 3 minutes later, only text conversation, no heavy work, simple answers using SONNET: 97%!

now after reset: exactly two interactions with 3 small files, simple text processing task, took ~4 minutes. 12% USAGE!!!

That really sucks.

now, less than 10 minutes later: 17%!! Simple, almost trivial work with Sonnet.

well i did normal test whit sonnet 4.6 to run commit, push, merge, commit, push and deploy this should take like 1% of usage and it took 97% of session work... something is really broken and cant do anything basically normal show me todo/tasks takes 17% usage just for that

Same issue I am facing. Any solution ?

same here, noticed that yesterday, token goes crasy fast like nosense!

I'm on Cursor Ultra using almost exclusively Claude Opus for couple months now. Last month used $460 over the entire month. I'm starting day 3 of this month and already used $80. Same work ~6x cost. I saw a single simple prompt jumped me 3%. Colleagues using CC report similar issues burning through $200 in a single day when previously nowhere near that.

Searching for answers I found this article covering the issue. I'm not affiliated with the site or the author. Unfortunately no solution.

There is no solution. The "solution" for me was to wait to the other day. Downgrade the code claude to stable version. I've tried opus 4.5 and sonnet 4.5 that evening with same results.
In the morning all started working normally. I think sometimes they have so much usage that they simply charge more tokens.
Since the morning all has been running normally but I'm monitoring the situation.

Well, "no solution" is incorrect. The solution is to simply stop using claude completely and use Codex. Problems solved :) It might not be a solution that anthropic would consider a "solution", but it is 100% a valid solution none the less.

i guess i was lucky today at morning/afternoon i got actually done the project work and debugs 4 hours on opus 4.6 and spend only 10% of weekly limit with 100% session usage. This was perfect lucky moment since i was struggling days of token usage and couldent do anything and today i actually could take back lost days and fixed many issues. Might be some fluke for my end this time, but we will see what next day brings again.

When this problem happened in January and got literally zero support from the company after weeks, I extended functionality of my mcp for Codex and have been using Codex Pro since as a primary, got rid of Max5x, and keep Claude Pro more for backup when I run out . I feel I get 10x more work done with Codex the way I work with Claude-Code to NOT need Max plans for it, and generally find it comparable to Claude minus the ability to spawn sub-agents. Occasionally when Codex gets stuck, I'll let Claude poke at it to fix (probably more just needs a fresh look with either), and go back to Codex.

Claude really needs to fix their support model if they expect to sell $200/mo plans to professionals that depend on it.

https://x.com/trq212/status/2037254607001559305?s=20

To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

https://x.com/trq212/status/2037254607001559305?s=20 > To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged. > During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

well i guess it is then good bye for claude. How developers can do there work if the work hours session tokens go faster -> answer is to use extra funds to get small task done... more money to company less for developers that uses it. this is a shame and this starts the shift where work is done outside office hours and soon the same effect is 24/7 more tokens taken less productivity done

https://x.com/trq212/status/2037254607001559305?s=20 > To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged. > During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

That X post is just a straight up lie. It absolutely DOES affect "weekly" limits and has been the case since Jan 1, 2026. I have LONG cut the cord from Anthropic COMPLETELY. I am now using exclusively Codex WITHOUT limit problems on their $20 subscription. I can work an ENTIRE 8hr day IN CODEX doing actual development work ALL DAY without running into 5hr nor weekly limits. For 5 days a week regardless of the times I do the work. Besides even if it DIDNT affect weekly (which we know it does, thats not up for debate, it's literally a known FACT), working for 15 minutes and then having to wait 5hrs to work another 15 mins is completely absurd.

well i guess it is then good bye for claude. How developers can do there work if the work hours session tokens go faster -> answer is to use extra funds to get small task done... more money to company less for developers that uses it. this is a shame and this starts the shift where work is done outside office hours and soon the same effect is 24/7 more tokens taken less productivity done

Working on global infrastructure for a long time now, you learn it's always go time somewhere. During downtime of seasonal work during december I hit claude pretty much 24/7 working on my own projects and it was solid, so much much so I bought into max. I could still almost tell when it'd slow down, like "oh, must be daytime in India." Then in general it fell apart the last week or so of december and became mostly unusable, and from that this thread.

Since this thread has been going I've run into this continually with Claude to the point I had to turn to alternatives to keep things moving with myself and my own business. At this point I watch it as a bit of a pulse. I've built my tooling for most platforms I use now but mostly I use Codex as a staple for what I need and do.

When I have to or try to do Claude, I feel like I get to work for 20 minutes and go into penalty hold. Boo.

GitHub Comment Draft — anthropics/claude-code #16856

Issue: [Bug] Excessive token usage in Claude Code 2.1.1 - 4x+ faster rate consumption
Status: OPEN, NOT LOCKED (55 comments)
Action: NEEDS AUDITOR REVIEW before posting.

---

Prepared Comment

The frustration in this thread is real, and I've been deep in the same problem.

After getting burned by a session that ate 20% of my weekly Max quota in under 30 minutes, I started parsing the JSONL transcripts that Claude Code writes to ~/.claude/projects/. The data tells a very specific story about where the tokens go.

Cache creation is the silent killer. In a typical session, cache_creation and cache_read tokens dwarf input/output. But the costs are asymmetric -- cache creation is expensive (same as input pricing), while cache reads are cheap. The problem: when context changes frequently (editing files, switching between tasks), the cache entries get recreated instead of reused. You pay the creation cost repeatedly without getting the read benefit.

Model selection compounds the issue. Every turn in my sessions used Opus, including trivial operations like reading a file or running ls. Sonnet handles these identically at 1/5 the cost. In one 37-turn session, 33 turns were simple single-tool operations that didn't need Opus at all. That's ~$4 in unnecessary spend in a single session.

If you want to see exactly where your tokens are going, you can parse the JSONL files yourself, or use a tool I built for this:

npx agent-forensics analyze ~/.claude/projects/-Users-you/

It breaks down: per-model costs, cache efficiency ratio, hot turns (the individual turns that ate the most tokens), waste patterns, and estimated savings. Runs locally -- nothing leaves your machine.

You can preview the report format without installing anything:

curl -s https://api.agentsconsultants.com/forensics/sample | jq .

The biggest practical fix I found: adding model routing hints to CLAUDE.md ("Use Sonnet for: file reads, grep, simple edits. Use Opus for: architecture, complex debugging"). That alone cut my next session's cost by ~60%.

Full disclosure: I built agent-forensics. It's MIT-licensed. The underlying problem is that Claude Code needs native cost visibility -- but until that ships, parsing your own logs is the best workaround.

GitHub Comment Draft — anthropics/claude-code #16856

Issue: [Bug] Excessive token usage in Claude Code 2.1.1 - 4x+ faster rate consumption
Status: OPEN, NOT LOCKED (55 comments)
Action: NEEDS AUDITOR REVIEW before posting.

---

Prepared Comment

The frustration in this thread is real, and I've been deep in the same problem.

After getting burned by a session that ate 20% of my weekly Max quota in under 30 minutes, I started parsing the JSONL transcripts that Claude Code writes to ~/.claude/projects/. The data tells a very specific story about where the tokens go.

Cache creation is the silent killer. In a typical session, cache_creation and cache_read tokens dwarf input/output. But the costs are asymmetric -- cache creation is expensive (same as input pricing), while cache reads are cheap. The problem: when context changes frequently (editing files, switching between tasks), the cache entries get recreated instead of reused. You pay the creation cost repeatedly without getting the read benefit.

Model selection compounds the issue. Every turn in my sessions used Opus, including trivial operations like reading a file or running ls. Sonnet handles these identically at 1/5 the cost. In one 37-turn session, 33 turns were simple single-tool operations that didn't need Opus at all. That's ~$4 in unnecessary spend in a single session.

If you want to see exactly where your tokens are going, you can parse the JSONL files yourself, or use a tool I built for this:

npx agent-forensics analyze ~/.claude/projects/-Users-you/

It breaks down: per-model costs, cache efficiency ratio, hot turns (the individual turns that ate the most tokens), waste patterns, and estimated savings. Runs locally -- nothing leaves your machine.

You can preview the report format without installing anything:

curl -s https://api.agentsconsultants.com/forensics/sample | jq .

The biggest practical fix I found: adding model routing hints to CLAUDE.md ("Use Sonnet for: file reads, grep, simple edits. Use Opus for: architecture, complex debugging"). That alone cut my next session's cost by ~60%.

Full disclosure: I built agent-forensics. It's MIT-licensed. The underlying problem is that Claude Code needs native cost visibility -- but until that ships, parsing your own logs is the best workaround.

# GitHub Comment Draft — anthropics/claude-code [#16856].

The session usage has absolutely nothing to do with claude code or how it handles things and everything to do with Anthropic has been fiddling with costs since Jan 1 2026. You can see the same problems in the web versions as well, it's not a claude code problem. They even recently posted and X comment about it, admitting they were messing with it during "peak hours". So all that token usage debugging you did was completely pointless as far as what your account usage shows since that's not the root cause :/

So the only "fix" is to simply stop paying anthropic and use something else like Codex. I have long cut the cord to Anthropic, Jan was when the subscription ran out and there are NO plans to renew it.

It's really crazy... This Thursday I asked Junie (Intellij IDEA's AI) with the Claude Opus 4.6 as an LLM some simple question, and it consumed 15$ not even being able to answer it (I only had 15$ left, so not sure how much more it would consume otherwise).

Well, one side benefit to them leaking source to claude-code, people can start debugging why it's broken easier since we can't get any support to do it.

Maybe related to this?

claude-code/blob/main/src/services/compact/autoCompact.ts#L68-L70:

“BQ 2026-03-10: 1,279 sessions had 50+ consecutive failures (up to 3,272) in a single session, wasting ~250K API calls/day globally.”

Not the first time it's happened. Also, claude code's src has ALWAYS been visible. The only thing this "leak" does is provide it in a non-minified format. For example: cat /opt/homebrew/lib/node_modules/@anthropic-ai/claude-code/dist/*.js gets you the EXACT same code that was "leaked", it's just minfied.

And the problems with usage are NOT claude code client related. It's simply Anthropic cooking the books (which they finally publicly admitted to in an X post after 3.5 months of complaints). These problems are visible even on the web chatbots. It just baffles the mind that anyone still even uses claude code. Get with the program, codex is the new king ;) And if you still want double model code reviews or whatever, kimi k2.5 is excellent at code reviews or even some of the local models if you have the GPUs to run them. At this point, Anthropic is just a dead dinosaur. There is no reviving it after the usage disaster.

This pattern continues with v2.1.89. Max 20, April 1 2026: 100% in ~70 min after reset. Getting worse, not better.

Full report: #41788
Related: #38335, #38239, #40790, #6457, #40895, #41055, #38345, #41174, #41550, #41617, #41663, #41779, #41802

I've been experiencing the same issue on Max 20 ($200/mo) — rate limit 100% exhausted in ~70 minutes.

After setting up a monitoring proxy using the official ANTHROPIC_BASE_URL env var, I identified two cache bugs as the root cause (#40524, #34629) and measured the impact: cache read ratio dropped to 4.3%, meaning ~20x token inflation per turn. After applying workarounds it stabilized at 89-99%.

Full analysis with per-request measured data, safe workarounds, and community references (including cc-cache-fix): https://github.com/ArkNill/claude-code-cache-analysis

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:

1. Update: claude update (or npm install -g @anthropic-ai/claude-code)
2. Pin the version: add "DISABLE_AUTOUPDATER": "1" to ~/.claude/settings.json env section
3. Avoid --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

The 4x consumption increase comes from context inflation — each turn re-sends the full conversation, and sessions silently accumulate bloat (progress ticks, file-history snapshots, duplicate CLAUDE.md injections, thinking blocks with signatures).

Cozempic v1.4.1 addresses this directly with 17 pruning strategies. On a real session: document-dedup alone saved 44% (duplicate CLAUDE.md re-injections), thinking-blocks saved 2-5%, and compact-summary-collapse saved 85-95% on compacted sessions.

The guard daemon auto-prunes continuously — install once, it runs on every session:

pip install cozempic && cozempic init

Happy to hear if anyone sees their consumption rate normalize after pruning.

And you are still using Claude code why? This should end all debate, it's not even a debate anymore, it's literally a fact, CC is no longer usable regardless of whether they fix their usage issues, cache issues, or claude code client issues. I've been saying this since Jan 1st and finally someone with some actual real reach woke up and realized it: https://www.youtube.com/watch?v=stZr6U_7S90

Saturday I tried switching up from Codex to Claude-Code as I was hammering my weekly quota, so figured I'd use Claude for a while to balance. I asked claude to review the codebase and task, and that we needed to engage in some feature troubleshooting. I got distracted and came back 20 minutes later, it was still churning, not even having done anything yet, I review my quota while its working, it went from 0 to 80% usage in that time. Finally it finishes to say it's ready to proceed with some recommended fixes. I tell it to proceed, and a minute later I get the "Out of Token, try again in 4.5 hours".

It's gotten simply unusable now. I guess I'll simply get another Chatgpt Pro account and use direnv to switch between them now until I can use it enough to justify the $200 plan, though that might come quick with this news yesterday from openai about api-based use for ALL codex now..

https://help.openai.com/en/articles/20001106-codex-rate-card

Sharing some token consumption data that might help anyone looking to optimize their workflows.

I instrumented my sessions to measure where tokens actually go:

| Source | Tokens per invocation | Frequency | % of session |
|--------|----------------------|-----------|-------------|
| Screenshot analysis | ~15,000 | 5-10x/session | 25-40% |
| DOM/page snapshots | ~114,000 | 2-5x/session | 30-45% |
| MCP tool definitions | ~8,000-70,000 | per session start | 10-20% |
| Retry loops (failed tool calls) | ~3,000-5,000 | 2-8x/session | 5-15% |

For the vision and DOM overhead specifically, I found that local preprocessing with Gemma 4 can extract the same semantic information at a fraction of the token cost:

• Screenshot → structured text: 15K → 400 tokens (97% reduction)
• DOM → semantic content: 114K → 500 tokens (99.6% reduction)

I packaged this into an MCP server that auto-detects your GPU (works from 8GB VRAM up) and picks the optimal Gemma 4 model. It also includes a retry loop detector that stops repeated identical tool calls at the 3rd attempt.

Interesting benchmark finding: the 2B model on 8GB VRAM was 2.7x faster than the 31B with comparable output quality for compression tasks.

For anyone interested in trying it: helix-agent (MIT, 330+ tests)

This doesn't replace the need for Anthropic to address the underlying consumption patterns, but it helps significantly with the vision/DOM overhead in the meantime.

Hey all, regarding the token usage problem, i've been building something for exactly this - https://github.com/Gerendo/engram
I hope it helps you. Trying to optimise and make it better on the go. It's basically a local semantic RAG index of your repo and gives Claude Code a search_<yourproject> MCP tool, so only relevant passages land in context instead of the whole codebase.

Would love feedback from people actually hitting the token ceiling - helps me making it better

I have the same problem since this monday and we are 06/23/2026 my usage is ridiculous And this is happening while I'm not even using it.
In 30 minutes, it went from 0% to 68% in two hours without any use.
Does anyone have a solution for me? I contacted Anthropic but haven't received a response.

I have the same problem since this monday and we are 06/23/2026 my usage is ridiculous And this is happening while I'm not even using it. In 30 minutes, it went from 0% to 68% in two hours without any use. Does anyone have a solution for me? I contacted Anthropic but haven't received a response.

i have same problem for few days, not doing anything it still eat up my token , i only 3 routine just sayhing hello , this is total stupid and oviously bug

@claude please do somthing, im not that rich

i hit the same after updating to 2.1.1, opus plan mode was eating tokens and i saw lock acquisition failed plus request aborted errors so my 5 hour limits got blown through about 4x faster. fwiw wozcode cut ur token spend ~50% w better caching, https://wozcode.com