What is Contexto?

There is a secret word. Every word in the dictionary has been pre-ranked by its semantic similarity to that word using AI embeddings — a technique that captures how closely words relate in meaning, not just spelling.

When you guess a word, you are shown its rank. A rank of 1 means you guessed the exact secret word. A rank of 500 means your word is in the ballpark. A rank of 10,000 means you are far away. The trail of guesses sorts itself so your best clues always rise to the top.

There is no letter information, no colour patterns based on position. Just pure meaning. It feels completely different from other word games — and once it clicks, it’s hard to put down.

The ranking is not a hand-written thesaurus and it is not a live call to a language model. It comes from a fixed set of word vectors — a table of 99,949 English words, each represented as a list of 50 numbers. Those numbers are the GloVe vectors (the 6-billion-token, 50-dimension release), trained by counting how often words appear near each other across a very large body of text.

The useful property of that table is that words used in similar contexts end up with similar numbers. Nothing in the training process was told that a doctor and a nurse are related; it simply observed that the two words turn up surrounded by the same other words, and so it placed them near each other.

To build a puzzle, the secret word’s vector is compared against all 99,949 others using cosine similarity— the cosine of the angle between two vectors, which gives a score from roughly −1 to 1 where higher means more alike. Every word gets a score, the whole list is sorted from highest to lowest, and your rank is simply your position in that sorted list. The secret word scores a perfect 1.0 against itself, which is why it sits at rank 1.

That whole ordering is computed once and cached, which is why a guess comes back instantly rather than after a round trip to a model. It also means the ranking is completely deterministic: the same guess against the same secret word always returns the same rank, for every player, on every device.

The single most counter-intuitive thing about this game is that rank is not linear in meaning. The distance between rank 1 and rank 100 is enormous. The distance between rank 9,000 and rank 9,100 is almost nothing.

That is a consequence of how similarity scores distribute. A handful of words are genuinely close to any given target, and then similarity falls off a cliff and the remaining tens of thousands of words are all roughly equally unrelated, separated by tiny fractions. Sorting that long flat tail still produces neat consecutive rank numbers, but those numbers are measuring differences of a thousandth of a point.

The practical consequence: moving from rank 12,000 to rank 6,000 tells you almost nothing — you may have got no closer in any meaningful sense. Moving from rank 300 to rank 150 is a genuinely large step. Treat improvements near the top as real signal and improvements deep in the list as noise, and the game gets considerably easier.

Every archive page shows this directly: alongside each past answer there is a small table of what the word at rank 51, 251, 1,003 and 10,001 actually was, with its similarity score, so you can see the curve flatten for that specific word.

The word table is ordered by frequency — the most common English words first. Secret words are drawn from a deliberately narrow slice of it: roughly positions 200 to 5,000.

The top of the list is skipped because it is almost entirely function words — the, and, was, of, it. Those make terrible puzzles. Function words appear beside every topic in the training text, so they end up sitting near everything in the vector space at once, which means no guess ever feels like it is getting warmer. They are excluded from being answers, from being category hints, and from being offered as “closer word” hints.

The long tail past position 5,000 is skipped for the opposite reason: it fills up with rare and obscure vocabulary that is unguessable in practice, and with words whose vectors are poorly trained precisely because they appear so rarely. A puzzle you cannot reason your way toward is not a hard puzzle, just a frustrating one.

Place names and proper nouns were removed from the dictionary as a separate pass. Plural forms are folded together with their singulars, so guessing cars when the answer is car is treated as the same word rather than as a near miss — and plurals are filtered out of the neighbour lists shown on archive pages, where they would otherwise crowd the top of every list with duplicates.

Which word lands on which date is derived from the date itself, so every player worldwide gets the same puzzle on the same day, and the archive can reconstruct any past day’s answer without storing a schedule.

There are three kinds of hint, and they are built in three completely different ways.

Every solve starts from 100 points. Each guess costs 1 point and each hint costs 3. The score is floored at 1, so an extremely long game still records a result rather than going negative.

The 3-point hint penalty is not arbitrary — it is an estimate of how many guesses a hint typically saves you. If a hint shortcuts about three guesses and costs about three points, then taking one is roughly score-neutral, and the decision becomes a genuine judgement call rather than an obvious yes or no. Price hints any lower and everyone takes all of them; price them higher and nobody ever does.

Ties break first on fewer guesses, then on who solved it earliest. The daily board shows a top ten and pins your own row beneath it if you placed outside, so the list stays readable without hiding anyone from their own result. The weekly board ranks by total wins across the week instead of points, which rewards showing up every day rather than one exceptional solve.

No guess limit. A cap would turn the game into a risk calculation about when to stop exploring. The interesting part is the search itself, so the search is unbounded.

No timer. Thinking time is the whole activity. Several of the best solves come from walking away and returning with a different angle on the category.

No letter feedback. The moment you know the third letter is an r, you stop reasoning about meaning and start pattern matching against a word list. Length and first letter exist as paid hints precisely so they stay a deliberate trade rather than a default.

No accounts required. The daily puzzle, unlimited mode and the leaderboard all work without signing in. Guest results are stored in your browser and listed under an assigned nickname.