How many words do you need to read a novel? We measured 14 classics

Original data, computed 2026-07-26. Method section expanded and the counting unit clarified on 26 September 2026; the numbers themselves are unchanged.

Ask this question in a language forum and you’ll get the folklore numbers: “3,000 words is conversational”, “8,000 to read novels”. So we counted. We ran fourteen public-domain classics in six languages through a word-frequency analysis and asked, for each book: if you knew the N most frequent words of the language, and no others, how much of this book’s running text would those words cover?

Two things to hold onto before reading the table. First, a “word” here is a lemma — a dictionary form, so was/were/been count once as be — ranked by frequency in modern general-purpose text. That is a finer unit than the word family used in most reading research (where happy, happiness and unhappy are one family), so our counts run well above the published ones. Second, coverage is not comprehension: it measures how many words you’d recognise, not how much you’d understand. Both points are unpacked below.

The numbers

Coverage = share of the running text (proper nouns excluded) made up of words from the top-N list. The last two columns show the list size where coverage crosses 95% and 98% — the two levels reading research associates with “followable with effort” and “comfortable”; see the coverage math for why. “—” means the level isn’t reached within the 50,000 most frequent lemmas of the modern language.

* Source edition uses pre-reform Portuguese spelling, which inflates these figures — see the finding on editions below. Names missing from the frequency list are excluded (knowing “Quijote” is not vocabulary); names that appear in it are counted — see step 3 of the method.

What the data says

1. A frequency list gets you less far than it sounds. The 2,000 most frequent lemmas cover 69–85% of a real classic’s running text: one unfamiliar word in every three to seven. The top 5,000 cover 76–92%. Past that point the curve flattens hard, because what remains is the long tail of words each book uses once or twice. Don’t read the 95%/98% columns as a study target, though. They are positions on a modern frequency list, and a nineteenth-century novel’s rarer words sit far down a list built from today’s web text and subtitles: countenance ranks around 24,000th among English lemmas, chaise around 33,000th. A reader who has met them in other novels doesn’t need the tens of thousands of modern words ranked above them. Reading research that counts word families puts 98% of a typical novel at roughly 8,000–9,000 families plus proper nouns (Nation, 2006) — the better figure for “how big a vocabulary”. What this table shows reliably is the shape: an ordinary course-sized vocabulary leaves a large gap, and that gap is made of rare words you mostly learn by reading.

2. The book matters more than the language. Within Spanish, Unamuno’s Niebla reaches 95% at ~12,900 lemmas while Don Quijote needs ~26,000 — twice as far down the same list, exactly the Golden-Age-Spanish warning from our Spanish guide. In German, Kafka’s Die Verwandlung (~17,900) undercuts Mann’s Buddenbrooks (~34,000) by nearly half — “Thomas Mann is a wall at B2”, now with a number on it.

3. Proust’s vocabulary is not the problem. The counterintuitive winner: Du côté de chez Swann needs fewer lemmas for 95% (~11,200) than The Three Musketeers (~17,700). Proust’s famous difficulty lives in his sentences, not his lexicon — the same lesson as Dostoevsky in our Russian guide: word counts measure one axis of difficulty, and syntax is another.

4. The edition can cost more than the author. Our two Portuguese classics look like the hardest books in the table — but much of that is spelling: the free editions use pre-reform Portuguese orthography, so thousands of tokens (elle, aquella, pharmacia) miss the modern frequency list entirely. It’s the same trap as German Fraktur and Russia’s pre-1918 letters, quantified: check the edition before you blame yourself.

5. The graded-reader gap is real and measurable. Top graded-reader series stop at ~3,000 headwords. The top 3,000 lemmas cover 72–88% of the classics here — followable in places, hard work in most — the gap the graded-readers article describes. (Headwords and lemmas are close enough units for this comparison; neither is a word family.)

Coverage is not comprehension

Every number above is lexical coverage: the share of running words a reader with a given vocabulary would recognise. It is a precondition for understanding, not a measure of it. Four things separate the two:

• The relationship is gradual, not a cliff. Studies that tested readers’ comprehension against the share of words they knew found it rising roughly steadily with coverage (Schmitt, Jiang & Grabe, 2011). 95% and 98% are commonly cited reference points — 98% for adequate unassisted reading of fiction (Hu & Nation, 2000), 95% as a minimal level (Laufer & Ravenhorst-Kalovski, 2010) — not switches where understanding turns on.
• Recognising a word isn’t knowing it in context. We count a token as covered if its lemma is on the list. That ignores other senses (fair the adjective vs fair the market), idioms and phrasal verbs, whose meaning isn’t the sum of known words.
• Syntax and background knowledge are separate loads. Finding 3 is the example: Proust’s words are more common than Dumas’s, and Proust is still the harder read.
• Real vocabularies aren’t frequency-ordered. The model assumes you know exactly the top N words and nothing else. Real learners have gaps in the top 2,000 and know plenty of rarer words from their interests, from cognates, or from earlier books.

So read the table as how much unfamiliar vocabulary each book puts in your way, comparable across books, and not as a prediction of how well any particular reader will understand it.

Method, step by step

1. Texts. The plain-text edition of each book from Project Gutenberg (ebook numbers below), with the Gutenberg header and licence boilerplate removed, normalised to Unicode NFC.
2. Tokens. A token is a run of letters, allowing internal apostrophes and hyphens. Hyphenated compounds are split and each part counted (drawing-room, va-et-vient). French and Italian tokens are also split at apostrophes (l’opéra → l + opéra), and 1–2-letter elision fragments missing from the list are treated as top-ranked function words. English possessive ’s is dropped. Digits are ignored.
3. Proper nouns. A capitalised token whose lowercase form has zero frequency in wordfreq is treated as a name and excluded from the counts. This only catches names the frequency list has never seen. Names that are on it stay in and are ranked like any other word: in Pride and Prejudice, Elizabeth ranks around 2,900th, Darcy around 14,500th and Bennet around 18,900th. Frequent characters’ names therefore push the 95% and 98% thresholds up somewhat, most in books with a small cast mentioned constantly.
4. Lemmas. Each remaining token is lemmatised with simplemma 1.2.0.
5. Ranking. For each language we take wordfreq’s (3.1.1) 50,000 most frequent word forms, lemmatise them the same way, sum the frequencies of all forms that share a lemma, and rank lemmas by that total. A token’s rank is its lemma’s rank, falling back to the surface form’s. For German, a word missing from the list is split into two known parts where possible (allowing a joining -s/-es) and given the rarer part’s rank. Anything still unmatched counts as beyond the list.
6. Coverage and thresholds. Coverage at N is the share of counted tokens with rank ≤ N. “Lemmas for 95%” is the rank of the token at the 95th percentile when all tokens are sorted by rank; if that token is beyond the list, the cell shows “—”.

The inputs and intermediate counts, so anyone can check the figures against their own run:

“Tokens” includes proper nouns; the two share columns are of all tokens and of non-name tokens respectively. Where “beyond the list” passes ~5% (archaic spelling, dialect, regional diminutives), the 95% threshold can’t be reached and shows as “—”.

Limitations. The frequency lists describe modern general text, not the literary language of the period, which pushes older books’ thresholds up. Gutenberg texts vary by edition and transcription. Lemmatisers make mistakes, more in some languages than others. And a single number per book hides how uneven difficulty is from chapter to chapter. For all these reasons, compare rows with each other rather than with figures from studies that count differently. The relative story (which books, which editions, which authors) is the robust part.

References

• Hu, M. & Nation, I.S.P. (2000). Unknown vocabulary density and reading comprehension. Reading in a Foreign Language, 13(1), 403–430.
• Laufer, B. & Ravenhorst-Kalovski, G.C. (2010). Lexical threshold revisited: Lexical text coverage, learners’ vocabulary size and reading comprehension. Reading in a Foreign Language, 22(1), 15–30.
• Nation, I.S.P. (2006). How large a vocabulary is needed for reading and listening? Canadian Modern Language Review, 63(1), 59–82.
• Schmitt, N., Jiang, X. & Grabe, W. (2011). The percentage of words known in a text and reading comprehension. The Modern Language Journal, 95(1), 26–43.
• Speer, R. wordfreq: word frequencies in many languages; github.com/rspeer/wordfreq. Barbaresi, A. simplemma; github.com/adbar/simplemma.

So what do you do with a 5,000-word vocabulary?

Read anyway, with the gap covered. The whole point of adaptive interlinear reading is that the 10–20% of words you don’t know arrive with their meaning already underneath, so a book with 84% raw coverage gives you something close to 98% of its words — while the rare words become familiar through use. It doesn’t untangle the syntax for you; that part is still reading. That’s what Glossa does with any book you upload and the classics in its catalog, including most of the books in this table. The free tier (two books, 10 pages in total) is enough to try it on a book “above your level”.

Start reading freeFree for 2 books and 10 pages of reading.

Related

• How to read books above your level
• Graded readers vs real books
• Interlinear books
• How to learn a language by reading