Skip to content

Glossary

Token

A token is a small chunk of text, usually a few characters or most of a word, that an AI model reads and writes one at a time, and the unit in which AI usage, limits and pricing are counted.

In plain terms

Models do not see words the way you do. They see text chopped into pieces, roughly three-quarters of a word each on average, so a thousand tokens is about 750 English words. When a plan says "200,000-token context window", that is the size of the pile of pieces it can hold at once.

01

Why it matters

Tokens are the meter on every AI tool you use. How much of a document you can paste in one go, the point at which a long conversation starts losing the thread, the moment a free plan stops answering, and the bill on a developer account are all counted in the same unit. Knowing roughly what a piece of work costs in tokens turns those limits from mysterious into predictable, and it is the difference between choosing a plan from its label and choosing it from what you actually do.

02

How it works

Before a model sees your text, a tokeniser cuts it into pieces drawn from a fixed vocabulary built while the model was trained. Common words usually survive whole. Rarer words break into fragments, so an unusual name or a technical term can cost three or four tokens where an everyday word costs one. Punctuation and spaces are counted too, and a leading space is normally folded into the token that follows it.

The model then works one token at a time. It reads the sequence so far, produces a score for every token that could plausibly come next, picks one, adds it to the end, and repeats. Nothing is composed a sentence at a time. That is why answers appear to type themselves out, and why a longer reply takes proportionally longer to arrive.

Both directions are counted. What you send is input and what comes back is output, and providers meter the two separately because generating text is the more expensive half. A conversation resends its own history on every turn so the model can follow the thread, so a long thread costs more per message than a short one even when your latest question is a single line.

Some models produce hidden working before their visible answer. That working is made of tokens too, and it is normally counted and charged even though you never see it, which is why the same question put to a reasoning model can cost several times what a direct answer costs. Where a provider exposes the setting, capping how much hidden working is allowed is a direct lever on both the bill and the wait.

Machine-readable formats cost far more than they look. Every bracket, quotation mark and comma is counted, and a repeated field name is paid for on every record, so the same information as structured output can run to several times what it costs as a sentence. That is worth knowing before a step is set to return a large table on every run.

Different model families tokenise differently. The same paragraph can come to different totals on two providers, which is why any figure you carry around is a rule of thumb rather than a measurement. English is the best-served case: text in languages the vocabulary covers less thoroughly breaks into more pieces, so the same meaning costs more.

One turn, and how it comes back round

One turn, and how it comes back roundThe sequence runs in one direction, start to finish. Your typed text enters a tokeniser, which cuts it into numbered pieces. The pieces enter the model as input. The model emits one output token at a time, each one appended to the sequence and fed back in as part of the next read. The finished pieces are joined back into readable text. Input and output are metered separately, and the whole conversation re-enters at the start at every turn.cutsas inputemitsjoined backthe wholeconversationre-enters atevery turnYour typed textTokeniserCUTS IT UPNumbered piecesMETERED AS INPUTThe modelOne output token at a timeAPPENDED, THEN READ AGAINReadable textMETERED AS OUTPUT
The sequence runs in one direction, start to finish. Your typed text enters a tokeniser, which cuts it into numbered pieces. The pieces enter the model as input. The model emits one output token at a time, each one appended to the sequence and fed back in as part of the next read. The finished pieces are joined back into readable text. Input and output are metered separately, and the whole conversation re-enters at the start at every turn.

One sentence, as the model sees it

One sentence, as the model sees itThe sentence is shown whole, then again cut into the pieces a tokeniser produces. Ordinary words survive intact, and the leading space belongs to the piece that follows it rather than standing on its own. Two words are the interesting ones. Neither is exotic, and neither has a single entry in the vocabulary, so each arrives as fragments joined back together. That is the whole lesson: this is unremarkable business English, nobody wrote it to break, and it breaks anyway. It is why counting words and counting tokens give different answers, and why the gap widens as the vocabulary gets less ordinary.THE TEXTOur finance team called theinternationalisation projectunjustifiable.CUT INTO 11 PIECES, FROM 71CHARACTERSOur·finance·team·called·the·internationalisation·project·unjustifiable.
The sentence is shown whole, then again cut into the pieces a tokeniser produces. Ordinary words survive intact, and the leading space belongs to the piece that follows it rather than standing on its own. Two words are the interesting ones. Neither is exotic, and neither has a single entry in the vocabulary, so each arrives as fragments joined back together. That is the whole lesson: this is unremarkable business English, nobody wrote it to break, and it breaks anyway. It is why counting words and counting tokens give different answers, and why the gap widens as the vocabulary gets less ordinary.

The same question, asked twice

The same question, asked twiceThe same question costs different amounts depending on where it is asked. In turn one of a fresh conversation the model reads only the question, perhaps thirty tokens, and writes its answer. In turn twenty the question is identical and the answer is the same length, but everything said so far is resent so the model can follow the thread, so it reads several thousand tokens instead of thirty. Input grew, output did not, and the bill follows input. This is why a long thread costs more per message than a short one, and why starting a fresh chat is the cheapest optimisation available.Turn 1 of a fresh chatYour question: about 30 tokens.Nothing before it to resend.The model reads 30 and writes200.Charged: 30 in, 200 out.Turn 20 of a long threadThe same question: still about30.Plus 19 turns of history behindit.The model reads 6,000 andwrites 200.Charged: 6,000 in, 200 out.The question did not change andthe answer did not get longer.What changed is how much had tobe read before answering, whichis why a long thread costs moreper message than a short one andwhy starting a fresh chat is thecheapest thing you can do.
The same question costs different amounts depending on where it is asked. In turn one of a fresh conversation the model reads only the question, perhaps thirty tokens, and writes its answer. In turn twenty the question is identical and the answer is the same length, but everything said so far is resent so the model can follow the thread, so it reads several thousand tokens instead of thirty. Input grew, output did not, and the bill follows input. This is why a long thread costs more per message than a short one, and why starting a fresh chat is the cheapest optimisation available.

Try it

Loading the interactive version...

03

Seen in the wild

  • Paste a long report into an assistant built for large volumes of source material and find where it stops taking the whole thing in one go. That ceiling is counted in tokens, not pages.

    Claude

  • Upload a spreadsheet and ask for a summary, then ask a follow-up with a second file attached. The second answer costs more, because everything attached is read again as part of the new request.

    ChatGPT

  • Send the same wording to two different model families through a single routing interface and compare what each one charges you for it. Identical words are not identical token counts.

    OpenRouter
04

Common misconceptions

People assume

A token is a word.

In fact

A token is whatever chunk the tokeniser's vocabulary happens to contain. Short common words are usually one token, longer or unusual ones split into several, and punctuation and spaces are counted as well. Roughly three-quarters of a word on average is a better working figure than one word.

People assume

Only what I type is counted.

In fact

The reply is counted too, and usually at a higher rate, because generating text costs the provider more than reading it. In a continuing conversation the earlier turns are resent each time, so the running total climbs faster than the length of your latest message suggests.

People assume

A bigger context window means the model reads all of it equally carefully.

In fact

The window is a capacity, not a promise of attention. Material buried in the middle of a very long input is more easily under-weighted than material at the start or the end, so a well-chosen extract often produces a better answer than the entire file.

05

Telling them apart

Tokens vs Words

Tokens

The unit the model and the meter both use. Fixed by the tokeniser's vocabulary, so it does not change with how you would read the text aloud. Punctuation, spaces and fragments of long words each count in their own right.

Words

The unit people use. Useful for briefing a writer and unreliable for predicting a bill, because 500 words of plain English and 500 words of code, names and figures are not the same amount of work for a model.

A limit written as a round number in the thousands is almost certainly tokens. A limit written for a person to plan against is probably words.

Tokens vs Characters

Tokens

Variable-length pieces the model was trained to recognise, so the count reflects how familiar the text is. Predictable enough to budget with once you know the rule of thumb, and the unit every meter actually uses.

Characters

A flat count of every symbol. Trivial to calculate and unrelated to how a model works: two passages with identical character counts can differ substantially once tokenised, because one is ordinary prose and the other is names and figures.

Character counts are worth using only as a rough proxy when nothing better is to hand. Anything that meters your usage is counting tokens.

06

Questions

How many tokens is 1,000 words?
Around 1,300 tokens for ordinary English prose, using the common rule of thumb that one token averages about three-quarters of a word. Text dense with names, numbers, code or non-English words runs higher, because those break into more pieces. Treat any figure you work out this way as a planning estimate rather than a measurement.
Why does my usage climb faster than the length of my questions?
Because a conversation resends its history. Every new turn carries the earlier messages so the model can follow the thread, so the input grows with each exchange even when your latest question is one line. Replies are counted as well, usually at a higher rate. Starting a fresh conversation for a new subject keeps the total down.
Do tokens work the same way on every AI tool?
No. Each model family uses its own tokeniser vocabulary, so the same paragraph can come to different totals on different providers. The general behaviour holds everywhere, with common words cheap and rare ones expensive, but exact counts do not carry across. Compare providers on the same sample text where the difference would matter.
Are spaces and punctuation counted as tokens?
Yes. A tokeniser processes everything in the text, so full stops, commas, line breaks and the spaces between words all consume part of the count. A leading space is usually folded into the token that follows it rather than counted on its own. Heavily formatted text therefore costs more than the same content written plainly.
Should I write shorter prompts to save tokens?
Only where the length is doing no work. Context that helps the model get the answer right first time is cheaper than a terse request that needs three rounds of correction. The savings worth chasing are structural: start a new conversation when the subject changes, and attach the relevant extract rather than the whole document.
How can I see how many tokens something is?
Several providers publish a counting tool, and developer interfaces report the totals for every request they handle. For everyday planning the rule of thumb is usually enough: treat 750 words of ordinary English as roughly a thousand tokens. Exact counts matter when you are building against an interface, rather than when you are drafting.
07

Key takeaways

  • A token is a chunk of text from a fixed vocabulary, averaging roughly three-quarters of an English word.
  • Usage limits, context windows and developer pricing are all counted in tokens, in both directions.
  • Replies are counted as well as requests, and a conversation resends its history on every turn.
  • Tokenisers differ between model families, so counts are estimates rather than portable figures.
  • The practical lever is what you send in, not how tersely you word it.
08

Related terms

09

Tools that use this

  • Claude

    Large context capacity, so its limits are where token counts become visible.

  • ChatGPT

    Attachments and long threads make the resend-the-history effect easy to observe.

  • OpenRouter

    One interface across many model families, so per-family token differences show up side by side.

Last checked July 2026

All glossary terms