The New York Times crossword · 1942–2025 · 2,457,203 answers
The crossword has never been written in English. It is written in whatever words happen to fit — and for fifty years those were lakes in Finland, Greek wine jugs and the geologist's word for a gravel ridge. Then, in the space of about a year, the Times threw the entire vocabulary out and replaced it with acronyms, brand names and television. The letters did not move at all.
Two lists, built by the same rule from opposite ends of the record. A word is retired if the Times printed it at least 50 times between 1942 and 1979 and has not printed it once since 2000. It is invented if the reverse holds. The twenty years in between are left unjudged, so neither list can be an artefact of where a boundary was drawn.
There is no overlap in character. The retired list is a reference shelf: a lake in Finland (ENARE), a Philippine island (SAMAR), an Austrian tributary of the Danube (ENNS), a ten-stringed Hebrew instrument (ASOR), the geological word for a gravel ridge (OSAR). The invented list is a channel guide and a form you have to fill in: SNL, SSN, EDU, TSA, LSAT, IMAC, IKEA.
Neither list is English. Each was a private vocabulary that a solver had to acquire separately from the language, and the clue attached to each word says so — the older ones are definitions, because you were expected not to know.
A vocabulary that was simply going out of fashion would fade: the words would drop away one at a time over decades. That is not what happened.
Eugene T. Maleska edited the Times crossword from 1977 until his death in August 1993. Will Shortz's first puzzle ran on 21 November 1993. Of the 56 retired words, the median was last printed in 1993. Of the 161 invented words, the median first appeared in 1994. Every one of the 56 was gone by 1999.
Fourteen of them — a quarter of the list — have their last appearance in 1993 itself, the year the editor's chair changed hands.
Here is the part that makes the swap look less like a change of taste and more like an engineering constraint.
A crossword answer has to cross other answers at every letter. That makes some letters cheap and others ruinous, and it pushes the whole vocabulary towards vowels and towards S, R, L, T, N. Crossword answers are 40.7% to 42.7% vowels — in every single year since 1942. English words, weighted by how often people actually use them, are 37.1%.
That gap has never closed. It was 42.1% in Farrar's first year and 40.7% in 2025: eighty-four years inside a two-point band. Compare the letter mix of the Farrar era with the Shortz era — two vocabularies that share almost none of their distinctive words — and they correlate at r = 0.997. Compare either with English and it falls to 0.96–0.97.
The puzzle did not get easier or harder. It swapped one source of vowel-heavy short vowel-heavy strings for another. ANOA, a wild ox, was last printed in 2004; ACAI, a berry, first appeared in 2010. Three vowels in four letters, either way.
One thing did change, and it is not the one people complain about.
Test every answer against ENABLE, the public-domain word-game lexicon — 172,823 entries, no proper nouns, no abbreviations, no phrases. From 1942 to 1993 the share of Times answers that were ordinary English words sat flat between 66% and 68% through three editors. It then fell for seventeen straight years to a floor of 53% in 2011, and has recovered only to 57%.
So the modern puzzle asks for fewer real words than the Maleska puzzle did, not more. What it asks for instead are names, initialisms and spoken fragments (ATAD, TADA, APOP). The complaint that the old puzzle was full of words nobody knew and the new one is full of things that are not words is, on this measure, correct in both directions.
The 20,000 most-used answers in the Times, each with its own eighty-four-year life. Type one in.
Every figure comes from the xd corpus, Saul Pwanson's archive of crossword puzzles, and specifically from its released clue file: 8.0 million answer-and-clue usages across 46 publications. This page uses only the New York Times rows — 2,457,203 usages of 205,829 distinct answers across 84 years — because the Times is the one publication in the corpus with continuous coverage under a single editorial lineage, which is what makes 1993 legible at all.
Answers are compared as bare strings of A–Z, upper-cased, with punctuation dropped, which is how the puzzle itself treats them: a grid does not distinguish E COLI from ECOLI.
1942 to 1950 is Sundays only. The corpus holds about 8,000 answers a year for those years against 31,000 from 1951 on, because the Times published only a Sunday puzzle until September 1950. Rates are unaffected — every figure here is a share — but a raw count from the 1940s is a count of Sundays.
The corpus is not the paper. It is a reconstruction, assembled from many hands, and it is not complete. 1978 is short by a quarter; 2012 to 2014 carry more rows than a year of puzzles contains. Shares absorb this; counts do not.
Editor years are approximate. Handovers happened mid-year — Shortz's first puzzle was 21 November 1993 — and the bands here are whole calendar years. The 1993 column therefore contains puzzles edited by both Maleska and Shortz, which makes the hinge slightly less sharp than the record, not more.
A trap worth naming. The first version of this analysis tried to measure the rise of abbreviations from the clues, by counting clues marked : Abbr. or for short. It showed abbreviations collapsing in the 1980s, which is false. Maleska flagged abbreviations by writing the clue itself in abbreviated form — Drivers' org. — rather than by appending a marker, so the marker count measures house style, not content. Nothing on this page is derived from clue formatting. The one clue-shaped fact here is independently visible: clues stopped ending in a full stop between 1967 and 1970, going from 98% to 5%.
The English baseline has a floor. Letter and vowel shares come from Norvig's word counts over the Google Web Trillion Word Corpus, weighted by frequency. That list stops at 333,333 words, and its tail is scanning noise, so it cannot say how rare ESNE is — only what the letters of ordinary English look like in bulk. No claim here depends on the tail.
Clues are quoted, not republished. One clue is shown per featured word, dated, as illustration. The full clue file belongs to the papers that printed it.
xd corpus — clue archive (xd-clues.zip, generated 28 April 2026) and puzzle metadata, Saul Pwanson / century-arcade. xd.saul.pw/data. Tooling MIT; clue and answer text remains the property of the publications that printed it.
ENABLE — Enhanced North American Benchmark Lexicon, 172,823 words, public domain, via dolph/dictionary.
Word counts — Peter Norvig, ngrams, derived from the Google Web Trillion Word Corpus (588 billion tokens over 333,333 words). Free to use.
Editor dates — the published record of the Times crossword: Margaret Farrar 1942–1969, Will Weng 1969–1977, Eugene T. Maleska 1977–1993, Will Shortz from 21 November 1993.