Quick Answer
Token limits are easier to plan around once you translate them into words. Using our measured average of 1,348 tokens per 1,000 words of English prose, 1,000 tokens holds about 740 words. That confirms the familiar "750 words per 1,000 tokens" rule for GPT models, with two important exceptions covered below.
Tokens to Words and Pages
| Tokens | Words (GPT) | Words (Claude estimate) | Single-spaced pages (GPT) |
|---|---|---|---|
| 100 | 74 | 55 to 65 | under 1 |
| 500 | 371 | 275 to 323 | under 1 |
| 1K | 742 | 550 to 645 | 1 |
| 4K | 2,967 | 2,198 to 2,580 | 6 |
| 8K | 5,935 | 4,396 to 5,161 | 12 |
| 32K | 23,739 | 17,584 to 20,642 | 47 |
| 128K | 94,955 | 70,337 to 82,570 | 190 |
| 200K | 148,368 | 109,902 to 129,016 | 297 |
| 1M | 741,840 | 549,511 to 645,078 | 1,484 |
Pages assume about 500 words per single-spaced page in 12pt Times New Roman. Double the page count for double spacing.
When Fewer Words Fit
- Claude models: Claude's tokenizer produces more tokens than OpenAI's for the same text, so each token holds fewer words. Plan on about 550 to 650 words per 1,000 tokens.
- Other languages: in our tests Turkish used 42% more tokens than English and Polish 54% more, so a 1,000-token budget holds far fewer words. Spanish and French were within 5% of English.
- Numbers, tables, and code: digits, symbols, and indentation break into many small tokens. Number-heavy English averaged about 1,480 tokens per 1,000 words.
Remember the Whole Request
A context window has to hold everything in a request: your instructions, earlier messages, pasted documents, and the model's reply. If a model has an 8K window and you want a 1,000-word answer, the question and any pasted material need to fit in what is left. For the largest windows available today, see what Claude's context window holds in words.