THIS GUY AUDITED 926 CLAUDE CODE SESSIONS AND FOUND MOST OF THE TOKEN WASTE WAS ON HIS SIDE everyone is blaming anthropic for the limits, so he decided to actually look at the data 858 sessions, 18,903 turns, and $1,619 estimated spend across 33 days here's what he found:... 1\ one default setting was burning 14,000 tokens per turn Claude Code loads the full JSON schema for every tool into context at session start. whether you use them or not. 20,000 tokens of tool definitions sitting there on every single turn. the fix: one line in your settings.json "ENABLE_TOOL_SEARCH": "true" context dropped from 45K to 20K instantly. across 858 sessions that one setting was wasting an estimated 264 million tokens 2\ cache expiry is the single biggest waste 54% of his turns came after a 5+ minute idle gap. every one of those turns re-processed the entire conversation at full price which caused a 10x cost jump you go grab coffee. come back 5 minutes later. type your next message. everything rebuilds from scratch. the context didn't change. you didn't change. the cache just expired. 12.3 million tokens wasted on idle gaps alone 3\ 42 skills loaded. 19 of them used twice or less across 858 sessions. every one of those skill schemas sat in context on every turn eating tokens for nothing. 4\ 1,122 redundant file reads where the same file was read 3+ times one session read the same file 33 times. he ALSO built a full token auditor dashboard that shows you exactly where your waste is coming from 19 charts, opens in your browser, free AND open sourceshow more

the 'user error' narrative is such a convenient shield for the devs. if someone has to audit 900+ sessions just to figure out why they're burning tokens, that's not a user problem, it's a UX failure. efficiency shouldn't be a puzzle the user has to solve just to avoid going broke

Jesus bro give some fkn details u clickbaity mfer

Isn’t this poor architectural choice by Anthropic though? There is no reason to load full context for tools and skills—some files Claude can read as needed (and there are several ways to orchestrate that but simple RAG will do). It is not the user’s job to find out.

"this isn't Anthropic's fault" *lists why it's Anthropic's fault

Which tool is that? I've used this one before, but that one looks different.

Here's the tool for this:

How are you reading this as a user problem? It's how Anthropic developed their product. You might even say it is nefariously programmed to consume your token allotment faster. Horrible fanboy take if you ask me.

This is a claude code design issue, not a user issue

This is one of the most useful Claude Code posts we’ve seen. Real data, not theory. The ENABLE_TOOL_SEARCH fix alone is worth the thread. Loading every tool schema on every turn is silent murder on your token budget. We hit the same bloat building Pelican’s multi-tool architecture and had to restructure how context loads for exactly this reason. The cache expiry finding is the one nobody talks about. You pause for five minutes to check a chart or read an article and your entire conversation rebuilds at full price. That 10x cost jump is real and it’s happening to everyone running long sessions. Two more areas worth auditing: redundant file reads aren’t just wasted tokens, they’re many chances for the model to subtly reinterpret your code differently across a session. And check for base64 encoded content persisting in context from file operations or image generation. That stuff sits there silently eating tokens across every subsequent turn.

loaded 42 skills and used two or less of them. felt this personally. i caught myself adding more and more context files to a client project last month when the actual fix was removing half of them. clarity is a subtractive process

where is the free AND open source link?

The redundant reads section hits hard. FWIW, jCodeMunch-MCP fixes that root issue—lightweight symbol retrieval so Claude doesn’t keep dumping full files..

the fact that stepping away to grab a coffee for 5 minutes is secretly costing you 10x more in token rebuilds is the most painful realization here

People are literally paying for tokens to ask Claude about the weather outside the window 2 feet behind them.

Most of the “Claude is expensive” take is just self own in disguise. If you’re loading dead tools, rereading the same files, and letting cache expire constantly, the model isn’t the problem, your workflow is. Thanks for sharing.

Seems like if we're talking about whom to blame, having default behaviors like enable tool search = false in Anthropic's own tool is not just Anthropic's fault, it's likely to be a dark pattern rather than an oversight.

ENABLE_TOOL_SEARCH is defaulted to true in Claude Code...

We built WOZCODE plugin to solve this. Claude is insanely wasteful with your tokens (even if you optimize to fix all these listed)

how is this the users fault? Anthropic is just shit, theyre making money on a shitcoded AI, this is scam

Plan to share repo?

Did he break down what specific patterns caused the most waste? I'm curious if it was context switching between projects or just verbose prompting that ate up tokens.

this is the kind of data-driven analysis the community needs. most token complaints are about the tool when the real issue is prompt hygiene. 926 sessions is a serious sample size and 'it was my fault' is an uncommon but valuable conclusion

tool search is on by default though...

@om_patel5 that's wild. didn't expect the waste to be on the user's side mostly. makes you think about how we interact with these models.

issue is lack of documentation guardrails and many more small things it burns lots of tokens to understand context that's why soon i will ship some good stuff. Even Claude says it loves it 🤭

Unpopular opinion. AI should know. Tell us how to be better at optimizing its token usage. Or better yet have a top layer that prevents us from wasting its time converts all the text to be better for Anthropic.

Right, so claude gave people a month refund just "because".

Wasting tokens and rug pulling are two completely different things. If i want to flush my tokens down the toilet then get out of my way. Don't keep rug pulling and squeezing your customers because your business model sucks

model routing is the real fix. haiku for grunt work, sonnet for code, opus only for architecture. most sessions don't need your most expensive model

I mean this just proves why its Anthropic's fault....looks like a profit increasing "scheme"..lol

nobody wants to admit that 'make the button bigger' doesn't need to be sent with the entire 40 file codebase attached

As far as I know, we cannot change cash expiration (5min). This one is really the biggest burner and there’s no way to fix it except respond within the five minute window or your SOL.

so the real limit isn't anthropic being stingy; it's us being too lazy to read the settings page before complaining on twitter

Claude Code was design to burn massive amount of tokens. More lightwieight harnesses produces fewer tokens and achieves better result.

This is actually great. A lot of people don't know about this.