Why Claude Code Uses 28K Tokens Before It Answers (vs Oh My Pi)
Claude Code sends about 28,000 tokens before answering even "1+1"; Oh My Pi sends about 10,000. Measured on one machine: what the system prompt and tool definitions cost, and two ways to stretch your usage limit.
What this video covers
- Same question (“what is 1+1”), same machine, four AI coding tools: Claude Code sent 27,659 tokens before answering, Codex 29,279, Oh My Pi 10,183, and the original Pi 1,164
- The load is the system prompt plus tool definitions, re-sent with every message: an instruction manual and a rulebook that ship again each time you speak
- Cutting 110 skill packs down to 12 changed almost nothing; removing my global rules file saved about 7,200 tokens per message (27,659 → 20,473); the remaining ~20,000 is Claude Code’s built-in tools and opening prompt, which users can’t remove
- Two things you can do today: keep your global rules file short, and check what’s taking up space (in Claude Code, type
/context) - The bigger saving is division of labor: expensive models think, cheap models do the work, and a fallback chain that stops and asks instead of quietly paying for a more expensive route
Chapters
- 0:00 Same question, very different bill
- 0:44 The numbers on the table
- 2:18 What’s inside the 28K
- 3:52 Why Oh My Pi only needs 10K
- 4:41 Who pays, and how
- 5:17 Two things you can do today
- 5:46 Expensive brains, cheap hands
- 6:19 When every backup fails
Transcript
Same question, very different bill
Ever feel like your Claude Code allowance runs out way too fast? I ran a little experiment: I asked Claude Code and Oh My Pi the exact same question — what is one plus one — then looked at what each one actually sent to the model.
A token is basically the unit AI bills by, roughly a chunk of a word. Claude Code burned about twenty-eight thousand tokens. Oh My Pi used about ten thousand. Nearly three times the gap, before anyone even said hello. This video covers three things: where the extra tokens go, how Oh My Pi gets away with less, and what you can do about it today. How does asking one plus one possibly cost twenty-eight thousand?
The numbers on the table
First, the numbers on the table. Claude Code, about twenty-eight thousand. Codex, about twenty-nine — slightly heavier. Oh My Pi, about ten. And the original Pi, under twelve hundred — remember that one, it matters later. And I haven’t asked anything real yet: this is just the pile each tool hands over before work starts. Every message you send gets stacked on top of it.
Why does the same question cost almost three times more? Because none of these AI coding tools have memory. You think you only said five words — write me a program. But before anything goes out, the tool packs a suitcase: what functions it can call, how each one works, the rules you gave it, what’s in this folder. In English that’s the system prompt plus tool definitions — plain words, an instruction manual plus a rulebook. Ask one question, the manual ships again — every single time, not one word shorter.
Think of it like a restaurant. You tell the waiter: one fried rice. A normal waiter knows the menu; you just name the dish. These tools make you read the entire menu out loud — plus the chef’s rulebook — before every single order. Order a hundred times a day, read the menu a hundred times. The menu reading costs more than the food.
That’s the gap between twenty-eight thousand and ten thousand. Not that one answers more carefully — they hand over different amounts of baggage before work begins. Remember two words: different baggage. Next, we open the suitcase piece by piece.
What’s inside the 28K
Second part of the bill: what’s inside the twenty-eight thousand. My biggest suspicion was that it was my own fault. I had a hundred and ten skill packs hooked into this tool — figured they had to be the burners. So I cleaned house: a hundred and ten down to twelve. Measured again.
Almost no change. Still about twenty-eight thousand. Ran it again, same number. Genuinely counterintuitive: you’d assume your own add-ons eat the budget, but the measurement says no. In Claude Code’s fixed cost per message, the skills line is basically zero.
Next I removed my global rules file — my personal rulebook, over nine thousand characters, sent word for word with every question. It dropped to just over twenty thousand. So that file costs about seventy-two hundred tokens every time I speak — mine, I own it. The remaining twenty thousand can’t be stripped: it’s the twenty-something built-in tools, each tool’s manual, and the opening speech Claude Code ships with. Users can’t change it; reinstall, it’s still there.
And it’s not just my machine. A team called systima measured it too: they intercepted Claude Code’s actual traffic and counted about thirty-three thousand at cold start — an older version, same ballpark. So it’s a general phenomenon, not my setup. And to be fair, this isn’t Claude Code being broken. It’s a product call: you’ll probably use these tools, so it brings all of them, every time. The thoughtfulness is just billed to you.
Why Oh My Pi only needs 10K
So how does Oh My Pi get away with ten thousand? Plain logic: carry fewer tools, write each manual shorter, read your rules from files you already have instead of shipping a new rulebook. But let me be straight with you: Oh My Pi is not the cheapest. The original Pi is way more extreme — under twelve hundred. Its author says publicly the whole opening plus every tool description comes in under a thousand tokens. That’s trimming to the bone. Oh My Pi goes batteries included — lots of features out of the box, several times heavier than the original, and the cost shows up in tokens. The comparison it actually wins is Claude Code and Codex: nearly the same features at about a third of the opening load.
Who pays, and how
Two kinds of people feel this bill differently. Pay per token — it’s literally money: ten thousand extra per message is real dollars over a month. On a subscription, like me, tokens burn quota: burn faster, hit the usage limit earlier; halve it and the same quota buys noticeably more questions. Plus a hidden layer: the AI can only hold so much. If the opening eats twenty or thirty thousand, less room is left for the actual conversation — chat long enough and it starts forgetting.
Two things you can do today
So two things you can do today. First, check your global rules file and stop stuffing it — mine cost seventy-two hundred per message; halving it saves three or four thousand. That’s the one valve you can turn yourself. Second, watch the token number your tool reports — in Claude Code, type slash context to see what’s taking up space. Don’t just ask whether the answer was right; look at how much you handed over to get it. Then the call — switch tools or trim rules — is yours.
Expensive brains, cheap hands
Third part — a bonus: the savings aren’t just tokens. The deeper saving is division of labor: expensive models think, cheap models lift. That deserves its own video, so just one line for now: the model doing the most real work on my machine is called S-W-E two — nearly ninety percent of the jobs — at about twenty dollars a month, the smallest line on my bill. Expensive brains for thinking, cheap hands for hauling, and the tool enforces the split itself.
When every backup fails
And when something breaks, it doesn’t go rogue. Every job type has a relay list written ahead of time: starter, backup, last resort. Real log on my machine — the morning of September twenty-ninth, two assistants were researching for me when the main model dropped offline. Three and a half minutes later the backup picked it up and finished — nobody called it, nobody asked me. Records of this died-and-auto-swapped kind: fifty-six. And the best part is the last rung: if every backup is dead too, the machine just stops, lays out who it tried and why it stopped, and waits for me to decide. It never quietly sneaks off to a pay-per-use channel to spend money. That rule I set myself — non-negotiable.
So back to you. If your Claude Code keeps hitting limits, it might not be that you’re asking too much — it might be that every sentence carries twenty-something thousand tokens of luggage.
What you’re looking at is two layers stacked. Same question: twenty-eight thousand versus ten — you save at the asking. Same job: the cheap model does the work — you save at the doing. And one last thing, to be clear: I’m not saying Claude Code is bad — it carries that stuff for reasons. Oh My Pi isn’t here to replace anything; it’s one more path. The choice of tool is still yours.
I measured every number in this video on my own machine, one run at a time. If it was useful, save it and send it to that coworker who keeps slamming into the usage limit. Now your turn: how many tokens does your AI tool carry before it says a word?
FAQ
How many tokens does Claude Code use before it answers?
On my machine, asking “what is 1+1” cost 27,659 tokens before Claude Code answered. Codex sent 29,279, Oh My Pi 10,183, and the original Pi 1,164. systima, intercepting Claude Code’s traffic on an older version, counted about 33,000 at cold start.
Why does Claude Code use so many tokens per message?
The tool re-sends its system prompt, every tool definition and your own rules file with each message — the models keep no memory between calls. In my test, skill packs added almost nothing, my global rules file added about 7,200 tokens, and the remaining ~20,000 is built in.
How can I make my Claude Code usage limit last longer?
Keep your global rules file short — it ships with every message — and check what’s taking up space by typing /context in Claude Code. Past that, the call is yours: switch to a lighter tool, or split the work so expensive models think and cheap models do the hauling.