Artem Horobchenko

Burned 236M Claude Code tokens in two weeks, then built a tool to fix it.
Now I audit yours and stop the leaks.

Full-stack dev building in public

[Artem Horobchenko @SSShken]

The token cost was only half the problem.

The other half nobody talks about: your AI has no memory of your project. You make real progress in one session, open a new one, and it has no idea what you built, why, or how it fits together. You spend the first 20 minutes just re-explaining your own project back to it.

For experienced devs it's annoying. For beginners building their first app, it's fatal, they lose the thread completely and the project falls apart.

The tools are incredible at writing code. They're terrible at helping you stay in control of what they wrote

[Artem Horobchenko @SSShken]

@NabilJB007 That overnight drop is rough. Funny thing, you working across 3 projects on Claude Code was probably what kept it low, more projects means more fresh contexts instead of one session growing all night

[Artem Horobchenko @SSShken]

@dagorenouf I love that!

[Artem Horobchenko @SSShken]

@pcshipp So before you showed login screen instantly?

[Artem Horobchenko @SSShken]

@marclou Wow every piece there is actual hoverable/clickable thing

[Artem Horobchenko @SSShken]

Two weeks ago I found out 99.9% of my Claude Code tokens were just the model re-reading its own context. My actual input: 0.03%.

So I changed how I work:
- one session = one task, then I close it
- no more marathon sessions that live for days
- when I need continuity, I write a short summary and start fresh instead of dragging the whole history

The difference is real. The sessions that used to quietly hit 100M+ don't happen anymore, because nothing is allowed to grow that big.

Turns out the fix was never a better plan or a cheaper model. It was just not letting context pile up.

[Artem Horobchenko @SSShken]

@heyandras Yeah, exactly i don’t get ppl who tell that new model is the best but come on it’s just came yesterday…

[Artem Horobchenko @SSShken]

@jackfriks @AnthropicAI I guess one week of setup and you're able to do magic on the phone

[Artem Horobchenko @SSShken]

@elonmusk I guess they're currently working out on an algorithm in a way so that even if it's open-sourced no one is able to manipulate it

[Artem Horobchenko @SSShken]

@AlexFinn I would spend at least week checking it out, remember what was with Kimi K3

[Artem Horobchenko @SSShken]

@gregisenberg Hope the amount of tokens is also half of Fable 5 ;L

[Artem Horobchenko @SSShken]

@claudeai Is that still half tokens of Fable 5?

[Artem Horobchenko @SSShken]

@phuctm97 How many tokens does it eat?

[Artem Horobchenko @SSShken]

@levelsio I mean you have just to balance everything, going to extremes isn't the best solution

[Artem Horobchenko @SSShken]

Which raises the obvious question: is the fix just discipline? Close sessions at task boundaries, start fresh, paste your own summary if you need continuity.

Curious if anyone here actually does that consistently, or if we all just let sessions run until they die.

[Artem Horobchenko @SSShken]

Someone posted their own token breakdown yesterday and the numbers lined up almost exactly with mine.

Him: 3B tokens, $3,586, across Claude Code, Codex, Cursor and Antigravity. 92.6% cache reads. Actual code output: 0.66%.

Me: Claude Code only. 99.9% cache reads. Output 0.4%. My own typed input was 0.03% of everything.

Two completely different setups, same shape. Almost nothing you pay for is the model producing work. It's the model re-reading context it already paid to write.

The gap between his 92.6% and my 99.9% is the interesting part. He runs four tools, so he's starting fresh contexts constantly. I live in one long Claude Code session until it collapses under its own weight. Spreading work across tools looks inefficient, but it accidentally keeps context small.

[Artem Horobchenko @SSShken]

@jxnlco It's funny, no matter what happens in life, we're always drawn back home

[Artem Horobchenko @SSShken]

@wickedguro Wait, bro, how are you supposed to make a video without a camera :D

[Artem Horobchenko @SSShken]

@phuctm97 "Chats can request starting a task on your computer"
That's what Claude, unfortunately, misses

[Artem Horobchenko @SSShken]

@AIHighlight It's weird for some reason, I don't see Codex in even half of them, even though everyone keeps saying it's the best...