Marusuke's Naruhodo Lab日本語Why did Claude Code use 10 times more tokens in one week than a study with 950 AIs?
Anthropic's big study put 950 AIs to work for 21 hours and made a new discovery. It used 210 million tokens.
But in the one week our Director spent building his own app, Claude Code on his Mac used 2.4 billion tokens.
Isn't the study's number way too small? Which one is right?
First guess
“Because the Director uses AI too much?”
He did use it a lot: 11,310 replies in one week. But the tenfold gap isn't about how much he used. 98% of the 2.4 billion was the AI re-reading.
Answer
Because the two numbers count different things: the Director's 2.4 billion includes the AI re-reading the conversation every time, and the study's 210 million doesn't.
Every time the AI replies, it re-reads the whole conversation from the start, so 98% of the Director's 2.4 billion tokens were re-reading.
Count both the same way, leaving out re-reading, and the Director's week comes to 48 million tokens, so the study used over four times as much.
Watch the video (5 min 21 sec)
Chapters (13)
Full transcript
0:00The question
NaruNine hundred fifty AIs, working for twenty-one hours, made a brand-new discovery. A huge research project by Anthropic.
NaruThe tokens it used: two hundred ten million.
NaruBut our Director spent one week building apps just for himself, and used...
NaruTwo point four billion tokens.
DirectorGo find out!
NaruHe says the research number is way too small. Which one is right? He wants me to find out.
NaruThe answer was hiding in a surprising habit of AI.
NaruMarusuke's Naruhodo Lab!
0:34The 950-AI research
NaruFirst, the news. Anthropic published this research on September 23rd.
NaruThey had 950 copies of Claude search a huge DNA database for unusual enzymes.
NaruThe AIs gathered over 200,000 candidates, narrowed them to 3,500, and wrote up the top 20 in reports people could read.
NaruAmong them was a system no one had noticed before, called ART. But what it actually does is still unknown.
1:06What is a token?
NaruAll that work, and just 210 million tokens. So, what exactly is a token?
NaruA token is a small piece of text that AI reads and writes in.
NaruAI chops text into little pieces like these. Every piece it reads, and every piece it writes, counts as a token.
NaruI've got it! Director, you're just using AI way too much!
DirectorRejected!
Naru...He says, show me the evidence.
1:34Logs on the Director's Mac
NaruAll right. Let's take a peek at the records on the Director's Mac.
NaruOn this Mac, an AI tool called Claude Code keeps a log of every exchange.
NaruFrom October 5th to the evening of the 11th, the AI wrote 11,310 replies. The busiest day alone used 690 million tokens.
1:56Quiz
NaruQuiz time!
NaruOut of those 2.4 billion tokens, how much do you think the AI actually wrote itself?
NaruThe answer is C: just half a percent. That's all.
2:1398% was re-reading
NaruNew reading was one and a half percent. And the remaining 98 percent was re-reading.
NaruThat's AI's surprising habit. Let's see why it happens, and how to cut it down.
2:28Why AI re-reads
NaruEvery time the AI writes a reply, it re-reads the whole conversation from the very beginning.
NaruThat's because the AI's mind resets to a blank slate after every reply. What remembers the conversation is the log kept outside the AI.
NaruIt's like re-reading your diary from page one every time you add a single line.
NaruOut of curiosity, I checked the very first reply in each conversation. In most of them, the AI had already read about 20,000 tokens.
NaruThat's because the instructions for its tools and the rules it follows are handed over at the start. It re-reads those every time, too.
NaruOn the Director's Mac, each reply re-read about 208,000 tokens on average.
Naru208,000 times 11,310 replies is about 2.35 billion. That's almost all of the 2.4 billion.
3:27The cache
NaruSo, is all that re-reading a waste?
NaruActually, it isn't. A system called a cache keeps what was already read, so going over it again is fast.
NaruAnd with the API, re-reading costs just one-twentieth of normal reading.
3:45New topic, new chat
NaruBut the longer a conversation gets, the more there is to re-read each time. In the longest one, a single reply read 690,000 tokens.
NaruSo when the topic changes, start a new conversation. That alone cuts re-reading way down.
4:03Leaving out re-reading
NaruNow, what if we leave the re-reading out?
NaruThe Director's seven days come to 48 million tokens.
NaruAnthropic's 210 million is more than four times that. The tables have turned!
4:17How the study counted
NaruSo how did the study count its 210 million? Their technical report spells it out.
NaruTokens re-read from the cache were left out. The study's 210 million doesn't include re-reading.
NaruSo the reversal we just saw, counted the same way, is the real answer.
4:38What we learned
NaruSo we can't say either number is wrong. Before comparing, you have to make sure they count the same thing.
NaruWhen you see a token count, first check whether re-reading is included. Naruhodo!
Director...Naruhodo.
NaruAnd the Director approved it!
4:57Your token count?
NaruIf you use AI, have you ever looked at your own token count? Tell us in the comments.
NaruThis has been Marusuke's Naruhodo Lab!
NaruIf you went "naruhodo!", please hit like and subscribe!
See it in pictures
More details
What is a token?
A token is a small piece of text that an AI reads or writes. The AI cuts text into these pieces, and both the pieces it reads and the pieces it writes are counted as tokens.
How were the Director's numbers counted?
On the Director's Mac, Claude Code, an AI coding tool, keeps a log of every exchange. From October 5 to the evening of October 11, 2026, the AI wrote 11,310 replies. The busiest day used 690 million tokens.
Each reply re-read 208,000 tokens on average. 208,000 times 11,310 is about 2.35 billion, almost all of the 2.4 billion.
Does even the first reply read a lot?
Yes. In most chats, even the first reply read about 20,000 tokens (the median of 188 chats). That's because instructions for the tools and the rules to follow are handed over at the start, and they are re-read with every reply too.
Is re-reading a waste?
No. A mechanism called caching keeps what has already been read, so it can be read quickly.
When you use the API, reading from the cache costs one twentieth of normal reading (Claude Opus 5.5 API pricing; this is not about subscription plans).
How can you cut down on re-reading?
The longer the conversation, the more each reply has to re-read. The longest conversation on the Director's Mac read 690,000 tokens in a single reply.
So when the topic changes, start a new chat. That alone cuts re-reading a lot.
What did Anthropic's study do?
The study was announced on September 23, 2026. 950 Claude AIs spent 21 hours searching a huge DNA database for unusual enzymes. They gathered more than 200,000 candidates, narrowed them to 3,500, and wrote up the final 20 as reports people can read.
Among them was a previously unknown system called “ART.” What it does is still not known.
TakeawayWhen you see a token count, first check whether it includes re-reading. Before comparing, make sure you're counting the same thing.
How we checked
- Anthropic research announcement (Sep 23, 2026)
- Anthropic technical report (token accounting on p. 30, "Campaign accounting")
- Anthropic, "Prompt caching" (cache pricing)
People and several AIs checked the video script against the text of the sources above and fixed misleading wording (October 11, 2026). Numbers are as of that date. AI wrote this page from that script and those sources, and a person checked it before publishing. If you spot a mistake, please email marusukelab@gmail.com.