The short answer: roughly 100,000 to 115,000 words for plain English prose. The honest answer is that it depends on what's actually in the text, and the gap between "best case" and "worst case" is bigger than most quick answers admit.
Where That Range Comes From
Running a sample of ordinary English prose — the kind of writing found in a novel, an essay, or a news article — through the same tokenizer used by GPT models (cl100k_base) produces a ratio of roughly 0.89 words per token. Applied to 128,000 tokens:
128,000 tokens × 0.89 words/token ≈ 114,000 words
Why the Answer Changes With Different Text
Running the same measurement on technical text — API documentation, code comments, content full of numbers, acronyms, and less common vocabulary — produces a noticeably lower ratio, around 0.68 words per token. The same 128,000 tokens works out to:
128,000 tokens × 0.68 words/token ≈ 87,000 words
That's a difference of roughly 27,000 words between the two estimates — purely from the style of writing, with the token count held constant. Technical vocabulary, unusual proper nouns, and non-English text all tend to break into more tokens per word than common English prose does, since the tokenizer has fewer single-token shortcuts for less frequent word patterns.
A More Reliable Approach
Since the ratio genuinely depends on the actual content, the more reliable method for anything that matters is measuring your specific text directly rather than relying on a general rule of thumb. Pasting in even a representative sample and checking its real token count gives an answer calibrated to your actual writing style, rather than a generic estimate that could be off by tens of thousands of words at this scale.
Get the real word-to-token ratio for your own text
Open the Context Window Estimator