How many words fit in a token budget?

For ordinary English, a handy rule is that one token is about three-quarters of a word.¹

So a budget of 1,000 tokens covers somewhere near 750 words.¹ That's a guess, though, not a measurement, and this page shows how far to trust it.

The makers' rule of thumb

OpenAI's help page, checked on 10 November 2026, gives three estimates for English text.¹

• One token is approximately four characters.
• One token is approximately three-quarters of a word.
• One hundred tokens are approximately 75 words.¹

The same page adds that these are estimates, not exact counts. Sentence and paragraph lengths vary.¹

A conversion table

The table below scales the "100 tokens is about 75 words" estimate up and down. The multiplying is mine, so treat each figure as approximate, in the same way the source does.¹

Anthropic's documentation says a number of its current models have a 1M-token context window.² The three-quarters rule comes from OpenAI, not Anthropic, so it can't simply be applied to a Claude window. Anthropic says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text.³ Fewer than 750,000 English words would therefore fit in a 1M-token window on those models, and I haven't tried to put a number on it.

Why pages and emails are harder

People often ask for pages or emails, not words. Neither maps cleanly. A page changes with font size and spacing, and an email might be three lines or three screens.

My advice is to convert in two steps. First turn the thing into words, using the word count your editing program shows. Then apply the rule. Because the rule is approximate, leave a good margin, and never plan to fill a limit to the last token.

When the rule breaks down

The estimates are for English. OpenAI says other languages can have different relationships between characters, words and tokens.¹

Anything that isn't plain prose may also behave differently. The same help page says a word can be divided into different tokens depending on its spelling, capitalisation and surrounding text.¹ So don't expect product codes or unusual names to follow the three-quarters rule neatly.

There's a second catch. Your own text isn't the only thing counted. Anthropic's context window page says the system prompt, tool definitions, images and documents all count toward the window.² So a 750-word message isn't the only thing using the budget.

Get a real count instead

Anthropic's token counting page also describes its count as an estimate, and says an actual request might differ by a small amount.⁴ Even so, a proper counter will beat mental arithmetic, because it uses the model's own method of splitting text.

Use the rule of thumb for planning. Use a counter when the number matters, such as when a long document sits near a limit. How to count tokens before you send shows how.

If you don't yet know what a token is, start with what is a token.