1.56
trips to Mordor
641,507 tokens for The Lord of the Rings trilogy without appendices.
estimateA million tokens. How much is that?
drag the slider
Books, chats, screams, rap and unicorns. All measured with one token ruler.
About this much fits into 1,000,000 tokens
9.18 ×
Harry Potter, book one
Selected item size: 108,991 tokens.
Choose what to compare
The slider changes the token budget. The buttons below only change the comparison.
Short texts
Songs & speech
Books
Big stories
Oddities
A model sees tokens, not words or pages. We counted familiar books, speech, short texts and a few odd examples to make any token budget easier to picture.
A token is a piece of text, not necessarily a whole word
A tokenizer splits text into short pieces and assigns each one a number. The model receives a sequence of those numbers and predicts the next token.
For common English text, a rough guide is about four characters per token. It is not a formula. Numbers, emoji, unusual words and formatting split differently.
Tokens fill the context window and affect request cost. GPT-5.6 currently has a 1.05M-token context window. Instructions, history, tools and the future answer all share that space. Current OpenAI models
This page uses o200k_base as one stable comparison ruler. It is not a promise of the exact count for every request to every model.
Open the OpenAI tokenizer
Count a complete model request exactly
- text
- The tokenizer quietly decides how much
- tokenizer
- The | tokenizer | quietly | decides | how | much
- token numbers
- 976 | 99665 | 58284 | 48145 | 1495 | 2009
- model
- 6 pieces entered the context
Try the real tokenizer
Paste text to see where token boundaries fall and which number each piece receives. The count runs locally in your browser with the o200k_base encoding.
The text stays on this device. It is not sent to a server.
Loading the local tokenizer…
0
tokens
0Tokens
0Characters
0Words
0Characters / token
The large text is the piece. The number below it identifies the token. A dot marks a space.
Enter text and its tokens will appear here.
The Museum of Token Oddities
A million tokens is roughly 2,222 book pages or 8.9 ordinary novels of 250 pages. Below are sixteen more comparisons. Books in the public domain were counted in full. Published word counts were used for books protected by copyright. Short texts and repeated symbols were counted exactly.
64%
of the Harry Potter series
Seven books take about 1,550,591 tokens. A million covers roughly two thirds of the series.
estimate1.30
copies of War and Peace
The measured English translation contains 768,982 tokens.
exact measurement27
trips down the rabbit hole
One trip means one copy of Alice's Adventures in Wonderland.
exact measurement175
U.S. Constitutions
5,725 tokens in the official transcript.
exact measurement455
performances of Rap God
Nearly 46 hours of uninterrupted rap.
estimate3.28
days of podcast
At 150 words per minute, that is 79 hours of talking.
typical length55
average days of speech
Based on a 2025 study estimate of 12,792 words per person per day.
estimate35
feature screenplays
Using a typical 110-page screenplay. Popcorn is outside the context.
typical length17%
of The Wheel of Time
The complete series is estimated at 6,043,852 tokens.
estimate66,667
messages saying “I’m leaving the subway…”
“I’m leaving the subway. I’ll be there in 5 minutes” takes 15 tokens.
exact measurement71,429
7-item shopping lists
Milk, bread, eggs, coffee, lemons, cheese and chocolate take 14 tokens.
exact measurement16,667
280-character posts
One example takes 60 tokens.
exact measurement333,333
unicorns
One 🦄 splits into three tokens. Checked on 1,000 unicorns.
exact measurement8,000,000
letter As in one scream
A sequence of 8,000 letter As took exactly 1,000 tokens.
exact measurement1,000,000
laughing faces
In a 1,000-repeat test, each 😂 took one token.
exact measurementFormatting is text too
Quotation marks, braces and field names are part of the text. The token count grows when the same word is placed inside quotes or JSON.
These are exact o200k_base counts. Tool schemas and message wrappers add more structure still, which is why OpenAI provides a model-specific Input Token Count API.
hello
1 token
hello24912
"hello"
3 tokens
"1hello24912"1
{"message":"hello"}
5 tokens
{"10848message3938":"7534hello24912"}18583
Everything we counted
The list is sorted by size. “Measured” means the available text was counted in full. “Estimated” means the figure comes from a published word count or a typical length.
| Fits in a million | How we got it |
|---|
What follows from this
Learn to understand AI model limits and work within them at Wai Combinator
Your text gets less than the full context limit
System instructions, chat history, tool results, images and the future answer share the context window. That always leaves less room for a pasted document.
The limit depends on the model
As of August 26, 2026, GPT-5.6 has a 1.05M-token context window. Other models have different limits, and those limits change over time.
Use the ruler for comparisons and the API for exact counts
o200k_base gives comparable results for every example on this page. For a complete request to a specific model, OpenAI recommends the Input Token Count API because message structure, tools, files and images also count.
How we counted
- Tokenizer
- Every comparison and the playground use the o200k_base encoding through gpt-tokenizer 4.0.0. It is one consistent ruler, not an exact count for every model.
- Texts
- English texts come from Project Gutenberg. Licence headers and footers were removed before counting. Russian texts come from ru.wikisource.org without tables and footnotes. When modern and pre-reform versions both existed, we counted only the modern one.
- Estimates
- For works protected by copyright, we did not use the full text. We multiplied a published word count by the measured tokens-per-word ratio for the relevant language. Those rows are marked as estimates.
- Language comparisons
- A translation is never an exact copy of the original. Translators build and divide sentences differently. The comparison shows an approximate scale, not an exact coefficient.
- Run the measurements
- The measurement script is in the repository at scripts/build-token-corpus.mjs. A complete request with messages, tools, files or images should be counted with the official Input Token Count API.