Tokens read so far0Every answer re-reads the whole conversation.
Why this session
The limitYou hit the ceiling at four in the afternoon, and you wait. Tokens are what filled it.
The invoiceTeam plan or API, the number on the bill is a number of tokens. Nothing else.
The answerA short, focused conversation answers better than a long, drifting one. Same tool, same model.
Saving tokens is not a finance topic. It is how you get better answers, faster.
Three things
What you leave with
One picture: what actually burns tokens when you talk to Claude. It is not what most people think.
Ten reflexes. Small, concrete, and you can start this afternoon.
Where the meter is, and the three signs that a conversation is getting expensive.
No prices, no formulas. The mechanism, and the habits that follow from it.
01
A token is attention you pay for
What it reads, what it writes, and why the two are not the same price.
Tokens
It does not read words. It reads tokens.
A token is a piece of a word. Roughly three quarters of a word, in English.
“Schneider Electric contract renewal”
Everything counts. Every word, every space, every comma. Yours, and its.
Two meters, not one
In and out are not the same price
InputWhat you send. It reads.
one token, one price
OutputWhat it writes. It thinks.
one token, several times that price
Ask for five lines, you pay for five lines. Ask for “a full report”, you pay for a full report you will not read.
Language
The same sentence, two languages
“Please send me the updated proposal before Friday.”
“Merci de m'envoyer la proposition mise à jour avant vendredi.”
Same meaning. French costs more, because the vocabulary was built mostly on English. Accents and long words cut into more pieces.
02
It re-reads everything, every time
The one mechanism behind every reflex in this session.
The mechanism
Message ten pays for the nine before
Your new messageThe only part that is actually new.
The document you attached at message fourRead again at five, six, seven… every time.
Everything beforeRe-read in full, to answer.
Orders of magnitude
The hidden heavyweights
A questiontwo lines, typed
≈ 30
One page of textan email, a memo
≈ 500
One screenshotof that same page
≈ 1,500
A ten-page PDFthe deck from last week
≈ 5,000
An eighty-page contract“just in case”
≈ 40,000
A spreadsheet, pasted wholeevery cell is text
≈ 100,000+
And each of these is read again at every message that follows.
Conversation A · message 1
Give me three arguments to convince a retail client to renew a two-year contract instead of one year.
Reads: your question. About thirty tokens.
Conversation B · message 52
Give me three arguments to convince a retail client to renew a two-year contract instead of one year.
Reads: your question, fifty-one earlier messages, two PDFs, a pasted spreadsheet. Tens of thousands of tokens.
Same question. Same answer. One of them cost a thousand times the other.
Live
Let me show you.
The same question, in a new conversation and at the bottom of an old one.
The answer is the same. What changed is everything the model had to read before writing it.
The idea behind every reflex
Saving tokens is not saying less.
It is not making it re-read what no longer serves.
03
Ten reflexes
Small, concrete, and none of them needs a setting. You can start this afternoon.
Reflexes
Before you start
1One conversation, one subjectNew subject, new conversation. The old one is not context, it is weight.
2Restart from a summaryWhen a conversation gets long, ask for a summary. Paste it into a new one. Ten lines replace fifty messages.
3Ask everything in one messageFive questions in one message are read once. Five messages are read fifteen times.
Reflexes
What you send
4The excerpt, not the fileThe three pages that matter, not the eighty-page contract. Read again at every message, remember.
5Text, not a screenshotIf the text exists, paste the text. A picture of a page costs three times the page, and it reads it worse.
Before you attach anything, one question: will I need this at message twenty?
Reflexes
What you ask for
6Say the length“In five lines.” “As a table.” “No introduction, no conclusion.” Output is the expensive side.
7Precise once beats vague three times“Do it better” replays the whole conversation, and guesses. Say what is wrong. Once.
8Fix small things yourselfA word, a date, a name: ten seconds of your time, or a full re-read and a full rewrite.
Reflexes
Which engine
9The model for the taskSort, reformulate, summarise, translate: the light model. Reason, decide, argue: the big one. Same conversation window, different price.
10Extended thinking, only when it earns itIt writes a long private draft before answering. You pay that draft. Brilliant on a hard problem, wasted on an email.
The model picker is not decoration. It is the biggest single lever you have.
Ten reflexes
One conversation, one subjectNew subject, new conversation.
Restart from a summaryTen lines replace fifty messages.
Ask everything in one messageRead once, not fifteen times.
The excerpt, not the fileWill I need this at message twenty?
Text, not a screenshotA picture costs three times the page.
Say the lengthFive lines. A table. No intro.
Precise once beats vague three timesSay what is wrong.
Fix small things yourselfA date is not a rewrite.
The model for the taskLight for routine, big for thinking.
Extended thinking, only when it earns itYou pay the private draft.
04
Set once, paid every message
Projects, instructions, memory, skills: what saves you from repeating yourself, and what it costs anyway.
Projects
Set once. Read at every message.
system
tools
your projectinstructions + twenty documents
the conversation
…room
0context limit →
What a project gives youInstructions and reference documents, posted once. No more pasting the house style in every conversation.
What it costsThe whole project sits in the box, in every conversation, at every message. Twenty documents “just in case” are twenty documents re-read each time.
A project holds what every conversation needs. Not the archive.
Three more places to set things once
Instructions, memory, skills
Custom instructionsWho you are, how you like answers. Written once. Short: it travels with every message.
MemoryWhat it already knows about you, you never repeat. Check it once in a while: stale memory is paid weight.
Skills and prompt librariesA good prompt, prepared once, replaces three messages of framing. The Natixis library is exactly that.
What it already knows, you never type again. And what you never type, it never re-reads.
Artifacts and documents
Edit. Do not regenerate.
Expensive
Rewrite the whole proposal with a stronger introduction.
It re-reads the proposal. It rewrites the twelve pages. You pay twelve pages of output, for one paragraph that changed.
Cheap
Change only the introduction. Make it stronger. Keep everything else as it is.
It re-reads the proposal. It rewrites one paragraph. Same result, a fraction of the cost.
Output is the expensive side. Name the part that changes.
05
Reading the meter
Where the number is, and the three signs a conversation is getting heavy.
Usage
Where the meter is
Session usagesettings → usage
Resets on a rolling window. Team admins see the whole team, per person.
It slows downEach answer takes longer. It is reading more before it writes.
“This conversation is long”The tool says it. Believe it: summary, new conversation.
You are waitingThe limit hit. The reflexes are what push that moment later in the day.
One sentence ties it all together
Every answer re-reads the whole conversation.
Saving tokens is not saying less. It is not making it re-read what no longer serves.
Before next time
Rewrite one conversation
Take
One real conversation of yours from this week. The longest one.
Bring
The same result, in three messages or fewer. And the reflex that did the work.
The ten reflexes fit on one page. It is yours to keep.