llmgauge

context window visualizer

What fits in a context window?

A context window is measured in tokens, which is hard to picture. Set a size and see what it actually holds — in words, pages, reading time, and code.

context size
200,000 tokens
10% of a 2M-token window
words
150,000
pages (≈500 wd)
300
reading time
630 min
lines of code
25,000
novels (≈90K wd)
1.67

Rules of thumb: 1 token ≈ 0.75 English words ≈ 4 characters; 500 words/page; 238 words/min; ~8 tokens/line of code. For exact per-model counts, use the token counter.

Tokens, not words

Models think in tokens — sub-word chunks averaging about four characters of English. "Tokenization" is why a 1,000-word email isn't 1,000 tokens but roughly 1,300, and why a block of minified JSON or a non-English script can be far denser. The window is the hard ceiling on prompt + context + output combined, so a long system prompt and a long answer both eat into it.

What the sizes mean in practice

A 128K window comfortably holds a long report or a medium codebase file-by-file. 200K reaches a short book. A 1M window spans a small repository or hundreds of pages — useful for whole-codebase reasoning or large document sets. But headroom isn't accuracy: retrieval quality often dips in the middle of very long contexts, so targeted context beats brute force when correctness matters.

The window is a ceiling, not a price

Picking a big-window model costs nothing extra by itself — you pay for the tokens you actually send, at the input rate. Fill a 1M window every request and the bill is large even on a cheap model. See the real number in the cost calculator.

Frequently asked questions

How many words is a token?

For typical English prose, one token is about ¾ of a word (≈4 characters). Code, numbers, and non-English text tokenize less efficiently, so the same character count uses more tokens. This tool uses the 0.75-words-per-token rule of thumb; for exact counts per model, use the token counter.

Does a bigger context window cost more?

Only for what you actually put in it. The window is a ceiling, not a charge — but every token you send is billed at the model's input rate. Filling a 1M-token window each call is expensive even when the per-token price is low. Model the real volume in the cost calculator.

Can a model use its whole context well?

Not always. Many models retrieve less reliably from the middle of a very long context (the 'lost in the middle' effect). A large window is useful headroom, but for accuracy-critical retrieval, well-targeted context usually beats dumping everything in.

Related