Reducing your token usage: ten practical methods
You hit your subscription limit by mid-afternoon, or your API bill doubles and you can’t see why. The cause is almost always the same, and it can be fixed.
What a token is, and why it matters
A language model doesn’t read words, it reads tokens: fragments of words. In English, a token is worth roughly three quarters of a word. Languages with accents and longer words, such as French or German, take slightly more of them to say the same thing. A page of dense text comes to somewhere between five hundred and eight hundred tokens, as an order of magnitude.
Tokens matter for three reasons.
- Price. If you use an API, you pay for every token the model reads and every token it writes. Written tokens generally cost several times more than read tokens.
- Limits. If you have a subscription, you don’t pay per token, but your usage is capped. The heavier your conversations, the sooner you reach the cap.
- Quality. A model drowning in pages of text it doesn’t need gives worse answers. It misses a detail, mixes up two versions, or forgets an instruction given at the start.
This third point is the most important: reducing your usage, done properly, improves the answers.
Where usage really comes from
The common belief is that usage depends on the length of your questions. It doesn’t. It depends on everything that is in the context at the moment the model replies, and that context is reread in full with each message.
Take a conversation where you pasted in a ten-page document at the start. On the first message, the model reads ten pages. On the second, it rereads those ten pages, plus your first exchange. On the twentieth message, it rereads the ten pages, plus the nineteen previous exchanges. The document has been read twenty times.
Here is what fills a context, from heaviest to lightest in most cases.
- Pasted or attached documents, reread on every turn.
- The conversation history, which grows with each message.
- Tool results: web pages visited, files read, search results.
- Standing instructions: custom instructions, descriptions of the connected tools.
- Your message, which is almost always the smallest share.
So when you are looking for savings, look first at what has been pasted in and what is lying around, not at how the question is worded.
Ten methods, from the most rewarding to the most fine-grained
1. Stop pasting whole documents
This is the biggest saving, by a long way. If your question is about one clause in a thirty-page contract, the model only needs that clause and what surrounds it.
There are two ways to get there. By hand, pasting only the relevant section. Or, far better, by connecting the assistant to a source where it can search and read only the passage it is interested in. See method 9.
2. Open a new conversation for each subject
A conversation that lasts all day drags its whole past along with each message. When you change subject, change conversation. If you need to keep a thread, ask for a ten-line summary of the old one and start the new one with it.
3. Write your standing instructions once, and keep them short
Custom instructions are added to every conversation. Three precise paragraphs are better than three pages that repeat the same thing. Keep what applies to everything, and move the rest into documents that the assistant will read when it needs them.
4. Give the full brief in the first message
Every round of corrections rereads the whole context. “Write an email” followed by five clarifications costs five turns. Say straight away who it is for, in what tone, at what length and with what expected result: you get the right answer in one turn.
5. Ask for answers of the right length
Written tokens are the most expensive and the slowest. If you want five points, ask for five points. If you only want the changed part of a text, say so, instead of letting the model rewrite the whole document with every tweak.
6. Choose the model to suit the task
Providers offer several models, from the lightest to the most powerful, with large differences in price. Rewording an email, sorting messages or extracting dates doesn’t call for the most expensive model. Keep that one for difficult reasoning.
7. Use extended thinking in measured doses
Extended thinking modes make the model produce intermediate reasoning, which counts towards usage. They are valuable on a complex problem and pointless for a simple task. Switch them on when the difficulty justifies it.
8. Make the most of caching
The APIs of the main providers offer caching: the part of the context that doesn’t change from one call to the next, such as long instructions, is billed at a much lower rate when it is reread. To benefit, put what is stable at the beginning and what varies at the end. If you are building an internal tool, this is often the setting that pays off most.
9. Have the assistant search instead of giving it everything
This is the fundamental change. Instead of putting your documents into the context, you connect the assistant to the place where they are kept. When it needs a piece of information, it runs a search, receives a few short excerpts with their location, and reads only the relevant part.
The difference is large on a long document.
| Approach | What the model reads | How often |
|---|---|---|
| Document pasted into the conversation | The whole document | With every message |
| File attached to a project | The whole document, or large portions of it | In every conversation that uses it |
| Search, then targeted reading | A few excerpts, then one section | Only when the question calls for it |
This is what the MCP protocol makes possible, as explained in MCP explained simply.
10. Organise and split up your documents
Method 9 only works if the documents lend themselves to it. An eighty-page catch-all file forces the assistant to read widely. Documents that each deal with one subject, with clear titles and subheadings, let it go straight to the point. We describe the method in Building a knowledge base for AI.
What doesn’t work
Some tips do the rounds a great deal and deliver very little.
- Dropping the pleasantries. “Hello” and “thank you” weigh a few tokens. A pasted document weighs thousands.
- Writing in telegraphic style. You save ten tokens and get an answer that misses the mark, which then takes three turns to correct.
- Translating everything into English. If your team works in another language, the saving exists but remains modest, and you lose precision when your documents and your readers are not English-speaking.
- Compressing documents into a summary. Useful for a conversation history, risky for a procedure or a contract: the missing detail is always the one you needed.
Measure before and after
You can only reduce properly what you measure.
With an API, every response shows the number of tokens read and written, and your provider’s console gives usage by day and by model. Look at the ratio of read tokens to written tokens: if you are reading a hundred times more than you write, your contexts are overloaded.
With a subscription, you don’t always have a precise counter. Notice at what point in the day you reach the limit, and in what kind of conversation. Command-line coding tools often display the size of the current context.
Run a simple trial: the same task, once with the document pasted in, once with the assistant connected to the source. Compare the usage and the quality of the answer.
In a team, the waste multiplies
On your own, pasting ten pages into every conversation is a bad habit. With ten of you, it is the same bad habit ten times over, on the same ten pages, every day. And each person pastes a slightly different version.
A shared source where each assistant reads the relevant passage solves both problems at once: usage, and the consistency of the answers. That is the subject of our guide Sharing context across your team’s AI assistants.
What Linkevio does about this
Linkevio is a shared knowledge base that your AI assistants read directly. It applies method 9 by design.
- The search covers the text of your documents and your imported files, and returns short excerpts with their location.
- The assistant then reads the part of the document it needs, not the whole document.
- Documents are organised by project and can be marked as authoritative, which saves the assistant from reading three versions to find one good one.
We won’t quote you a savings percentage: it depends on the length of your documents and on what you used to paste in. The longer your documents and the more often you used them, the bigger the gap. The free plan lets you check this on your own cases: create your workspace and run the trial described above.
Frequently asked questions
Why do I reach my subscription limit so quickly?
Almost always because of long conversations that contain large documents. Every message rereads everything. Open a new conversation for each subject and avoid pasting whole documents: that is what makes the most difference.
Does a shorter question cost less?
Barely. Your message is the smallest share of the context. What counts is what comes with it: documents, history, tool results.
Doesn’t connecting tools through MCP use tokens?
Yes, a little: each tool’s description takes up room. That is why it is better to switch on only the connectors you use. But a connector that returns excerpts instead of whole documents saves far more than it costs.
Does caching apply to my subscription?
Caching is first and foremost an API mechanism, which concerns people who build things. In a subscription, optimisations are handled by the provider; you don’t need to deal with them.
Are files attached to a project reread every time?
It depends on the assistant and on the size of the files: some load everything, others fetch portions. Either way, they are copies that you have to remember to update.
Does cutting tokens make the answers worse?
No, if you remove what is unnecessary and keep what is needed. A shorter, better-targeted context generally gives better answers. What makes them worse is removing information the model needed: hence the value of letting it search for itself.