Contexto Answer Today (September 13, 2026)

Contexto hides one secret word and gives you exactly one piece of information per guess: where that guess ranks in semantic similarity to the answer. A rank of 250 means your word is closer to the answer than 249 others and farther than most of the dictionary. Rank 1 is the answer itself. That's the entire interface. No letters, no colored tiles, just a number.

The beginner trap is treating small numbers as praise. Type dog, see 3,000, and it feels like progress. It is not. The number is a coordinate, not a compliment, and if you do not move toward it you stay lost.

The engine behind all of it is a word-embedding model trained on an enormous pile of text. Words get mapped to vectors, and similarity is measured by how close those vectors sit. That's why synonyms rank well but so do words that merely appear in the same contexts. A good guess doesn't have to mean the same thing; it has to live near the answer in the space the model learned.

Once that clicks, you stop solving a crossword and start reading a map. Every rank becomes a reading on that map, and the fastest route to the answer is triangulation: plant three or four anchors around the target and walk inward.

Opening with the fanciest word you can summon assumes Contexto rewards cleverness. It doesn't. The model rewards semantic centrality, and words like house, water, time, and people sit in the densest parts of the space. They're boring, but they're useful distance probes even when they're far from the answer.

A clever word like serendipity sits in a sparse region. If it ranks 15,000, you've learned almost nothing about which direction to walk, because there simply aren't many words nearby to compare against. A boring word like street ranking 4,000 tells you the answer lives in a populated neighborhood with plenty of reachable words, and that is something you can act on.

The solver on this page leans on exactly that principle. It tracks the ranks of every guess, models the semantic neighborhood, and suggests the word most likely to shrink the distance fastest. It exists to stop players from burning turns on impressive-sounding dead ends.

One low rank tells you the answer is nearby but not where. Two low ranks in the same family confirm the lane. Three low ranks that bracket the answer from different angles, an emotion, an action, and an object all under 500, hand you the answer within a couple more guesses.

A common early mistake is choosing anchors that all point the same way. If a first guess ranks 800 and a second near-synonym ranks 900, that confirms the lane but reveals nothing new. The right second guess probes an adjacent lane, a related but different word, to see whether the answer sits between them.

The endgame is a shrinking circle. Once guesses start ranking under 100, the phase shifts from exploring to converging: near-synonyms of the best word, then near-synonyms of those. It is mechanical, and that is the point.

Contexto rewards pattern recognition more than raw vocabulary, and patterns are learnable. Replay old puzzles through the solver and study the path from first guess to answer: which guesses move you into the right lane, and which one wastes a turn. The wasted turns are almost always clever words in sparse regions.

Choose the opening deliberately, every time. The gap between opening with house and opening with serendipity is the gap between a 15-guess solve and a 40-guess solve. The opening sets the semantic anchor for everything that follows.

Treat every loss as a map of the model's quirks. Contexto answers are occasionally surprising, and a word that ranks 50 may not mean what you assumed. The model's associations are the ground truth, and the faster you learn them, the faster you solve.

Contexto answers skew toward common words, not exotic vocabulary, because the ranking model is trained on how people actually write. The answer is far more likely to be a word like current, office, or partner than equanimity. When your low-ranking guesses are all uncommon words, the answer is probably a common neighbor you're walking straight past.

Nouns and verbs behave differently in the ranking, and knowing which you're chasing changes everything. Nouns cluster tightly; the model keeps bank, money, and loan close together. Verbs spread across many contexts. If the answer is a noun, the lane strategy works fast. If it's a verb, the ranks stay stubbornly high for longer, and the solver's suggestions become more reliable than raw verb guesses.

Adjectives are the trickiest lane because they pair with everything. A guess like happy can rank well whether the answer is cheerful, satisfied, or thrilled, so a good adjective rank tells you the feeling but not the word. The solver handles this by probing several adjective anchors before converging. Copy the habit: one emotion word, one action word, one object word, then read the map.

Behind the rank number sits a vector: a list of a few hundred numbers that summarize where a word sits in the model's learned space. Every word in the dictionary gets one of these vectors, and the list is built by training a neural network on a huge slice of ordinary text. The training task is simple to describe and brutal to scale: given a word, predict the words that tend to appear near it. The list of numbers is the side effect, not the goal, but it turns out the side effect is exactly what Contexto needs.

Two vectors that point in similar directions belong to words that show up in similar contexts. King and queen sit near each other because they share neighbors like royal, throne, crown, and reign. Walk and run sit near each other because they share neighbors like fast, slow, race, and distance. The model never sees a dictionary, and it never reads a thesaurus. The neighborhoods come from raw text, which is why some pairs feel obvious and some feel strange: the model is following statistical co-occurrence, not editorial definitions.

Dimensionality is the part that surprises people. The vectors are not three-dimensional points on a graph. They are points in a high-dimensional space, typically a few hundred axes, where each axis captures some latent feature the model invented during training. Nobody named those features. They are not "nounness" or "verbness" or "kitchen-ness." They are statistical patterns the network found useful for the prediction task, and the only honest description of any one axis is a fuzzy label like "this dimension helps separate indoor nouns from outdoor nouns." You cannot read a vector by hand, but you can read the distances between vectors, and Contexto does that for every guess.

Contexto also adds one trick on top of the raw vectors: it does not use raw cosine similarity. It uses a game-tuned similarity function that reweights the dimensions, which is why some pairs land much closer than a vanilla embedding model would put them. The exact recipe is the developer's, but the effect is visible to anyone who plays a few rounds. Words that are not synonyms still get pulled into the same neighborhood when they share a topic, a domain, or a usage pattern. The space is not a thesaurus, and it is not a Wikipedia category tree. It is a learned geometry that has been tweaked for the puzzle.

The practical takeaway: the rank you see is the model's opinion of how close your guess sits to the answer in this tuned space, and the model's opinion is shaped entirely by what the training text looked like. That is why some neighborhoods feel intuitive and others feel arbitrary. You are reading the geometry of a large text corpus, not a definition.

Polysemy is the first thing that breaks player intuition. A word like bank has two clean senses: a place for money, and the side of a river. The model has to pick one vector, and it picks the average, weighted by how often each sense appears in the training text. The money sense usually wins, because financial writing dominates, and that is why bank, money, and account sit close together while bank, river, and water sit farther apart. The same effect hits words like bat, match, set, light, and almost any common noun that has shifted meanings over time.

Topic contamination is the second surprise. Two words that share a topic can rank as close neighbors even when they are not synonyms at all, because the model picks up on shared context. Doctor and hospital are not synonyms, but they share a topic, so they sit near each other. Cat and dog are not synonyms either, but they share a topic, so the model groups them. That is the reason Contexto answers often belong to a topic the player can name, even when the player cannot name a synonym. The answer is not the same word as your guess. It is a word from the same neighborhood the model has decided is the right one.

Frequency warps the geometry in a quieter way. Common words have stable, well-trained vectors, because the model has seen them in thousands of contexts. Rare words have noisy vectors, because the model has only seen them a few hundred times. The consequence is that a rare-word guess gets a rank that is partly a measurement and partly noise. A rank of 350 for a common word is a real 350. A rank of 350 for a rare word is a 350 with a margin of error that might be a hundred in either direction. Guess rare words only when the surrounding ranks tell you the answer is itself a rare word.

Negation is the third quirk worth knowing. The model has no concept of not, no, or without as operators. Anton and antonym are not the opposite of cat and dog in the vector space. The model knows that not and cat appear in similar contexts to no and cat, which is not the same thing. If the answer is a negative of a common word, the only way to find it is to guess the positive and then walk the neighborhood by meaning, not by negation. There is no shortcut the model offers here, and players who try to guess with not stuck on the front waste turns.

Morphology helps more than people expect. The model treats walks, walked, walking, and walk as neighbors, because the surrounding text overlaps heavily. That is why verb tense barely matters in Contexto: guess the base form when you can, because the model will forgive the mismatch. The same applies to plurals. Guesses and guess land close, and so do cat and cats. The opposite is true of truly different words that share a stem. Help and helpful are not synonyms, and the model knows it. The overlap is real but shallow, and ranks reflect that.

Give the puzzle an honest run before checking today's Contexto answer, and hold off on the reveal until later in the day. This isn't about discipline, it's practical: the moment the answer appears, the puzzle is over and there's nothing left to learn from it. The reveal card at the top of this page is there for the days you're stuck or in a hurry, and it's confirmed against the official puzzle rather than guessed.

What the daily reveals shows, more than any single word, is the shape of the model's sense of meaning. Each day's answer shows which words the model considers close, and reviewing those, even briefly, sharpens your intuition for the next one. There are no letter clues and no guess limit; the only thing standing between you and the answer is how well you can read the map.