Semantle: the website that makes you guess by meaning

Semantle is a daily word game that hides one secret word and scores every guess by how close it sits in meaning, not in spelling. It is worth a look because it takes the familiar once-a-day puzzle and quietly swaps the rules of thumb everyone learned from Wordle for something stranger: a cold number that measures whether two words tend to appear in the same company.

There is no grid, no colored tiles, and no six-guess limit. A player types a word, and the site returns a similarity score. Warmer numbers mean the guess lives near the answer in a map of language built by a machine. Colder numbers mean it does not. The whole interface is a text box and a running list, which is exactly the kind of restraint that makes a single-purpose website easy to recommend.

In brief

• Semantle reveals one secret word per day and ranks each guess by semantic similarity rather than letters, so spelling and length give no hints.
• The score comes from word2vec, a model that represents each word as a 300-number vector learned from a large body of Google News text.
• It was built by David Turner in early 2022 as a deliberately harder answer to Wordle, with unlimited guesses and no clock.
• Easier spin-offs exist, including Semantle Junior, a Spanish edition, and a Teams mode for playing together.

What Semantle actually asks of a player

The task sounds simple and is not. Each day the site picks one secret English word. A player guesses words, one at a time, and the game answers each guess with a similarity figure. According to the official Semantle FAQ, the objective is to guess that secret word, and every guess must be a single word. Meaning is the only currency. A guess can share no letters with the answer and still score well if the two words keep similar company.

That inversion is the point. In most word games the surface of the word matters: its length, its letters, their order. Here the surface is irrelevant. The word "ocean" might score highly for a secret word of "sea" while "oceania" sits nowhere near it. The player is not spelling toward an answer. The player is reasoning about which concepts cluster together, then testing that hunch one noun or verb at a time.

The feedback loop is blunt by design. There are no yellow or green squares to narrow a field. There is a number, and a player must decide what it implies. A rise might mean the last guess moved in the right direction, or it might be noise from a quirk in the underlying data. Interpreting that single figure is the entire game.

How the scoring works: meaning, not spelling

The score is not a human judgment. It comes from word2vec, a technique Google researchers introduced in 2013 to represent words as dense lists of numbers. The original paper by Tomas Mikolov and colleagues, Efficient Estimation of Word Representations in Vector Space, described training a model so that words used in similar contexts end up with similar vectors. Words are not defined. They are located.

The idea landed because the vectors captured relationships, not just neighbors. Word2vec became famous for a kind of arithmetic in which operations on word vectors produced sensible results, the canonical example being that the vector for king, minus man, plus woman, lands near queen. Semantle never shows that machinery, but it is the same machinery underneath. Every score a player reads is a small consequence of words having been turned into geometry.

Semantle uses one well-known set of these vectors. The FAQ names the GoogleNews-vectors-negative300 dataset from late 2021, and the Wikipedia entry on Semantle notes that each word is represented by a 300-dimensional vector. Every word becomes a point in a 300-axis space. To compare two words, the game measures the cosine of the angle between their vectors, a standard way to ask how aligned two directions are regardless of their length.

The scale of that training is part of why the scores feel authoritative. The word2vec project page from Google describes its published GoogleNews model as covering roughly three million words and phrases, with vectors learned from about one hundred billion words of news text. Semantle inherits all of that statistical memory. When the game says two words are close, it is reporting a pattern distilled from more reading than any person could do in a lifetime.

This is why the game can feel uncanny. The model learned its sense of nearness from how words actually co-occur across a vast pile of news text, not from a dictionary. So "coffee" sits near "tea" and "espresso," but also near "morning" and "cup," because that is the company the word keeps in print. A player is, in effect, navigating the statistical habits of written English.

The 1000-word proximity ladder

Raw similarity numbers are hard to read on their own, so Semantle adds a ladder. The FAQ explains that once a guess lands within the thousand words closest to the secret word, the site reports its position in a proximity column. The nearest word to the answer ranks near the top of that scale, and the ranking shrinks quickly as guesses drift outward.

That rank is the game's one act of mercy. A similarity of, say, a few percent tells a player almost nothing. A rank of "within the top 1000, position 50" tells them they are circling the target and should press on nearby ideas. Most guesses never enter the top thousand at all, which is how a player can spend dozens of attempts learning only that an entire region of meaning is cold.

Where it came from: a reaction to Wordle

Semantle did not appear in a vacuum. It arrived in early 2022, at the height of the Wordle craze, and reads as a direct response to it. The Wikipedia entry credits the game to the developer David Turner and places its creation in January 2022, during the stretch when a single daily puzzle had become a shared social ritual. The official FAQ confirms the credit, noting that Semantle was created by David Turner.

The contrast with Wordle is instructive. Wordle, built by the engineer Josh Wardle and released publicly in October 2021, gives six tries at a five-letter word and tidy color feedback. Its popularity was large enough that, as Wikipedia records, The New York Times acquired it on 31 January 2022 for a low seven-figure sum. Wordle is finite and forgiving. A player either solves it in six or does not, and either way the day is over in minutes.

Semantle took the daily format and removed the guardrails. There is no six-guess cap and no five-letter constraint. The puzzle refreshes on a schedule the FAQ gives as 8 PM New York time, and until then a player can guess as many words as patience allows. The game kept the ritual and discarded the mercy.

Why it feels unsolvably hard

Players describe Semantle as infuriating, and the reasons are structural rather than accidental. The Hacker News discussion of the game surfaces the core complaint: a single distance number gives little direction. When a guess is far from the answer, small wobbles in that number are often noise, not signal, so a player can chase a rising figure straight into a dead end.

A subtler problem lives in the data. Because the vectors were trained largely on news articles, the model's sense of nearness carries the fingerprints of news coverage. Words that share headlines end up close even when a dictionary would keep them apart. The same Hacker News thread raises this point, and it explains the occasional guess that scores strangely high for reasons that feel more journalistic than linguistic.

Then there is the shape of the search. Human intuition about meaning is rich but low-resolution. The model's space has 300 axes, and the path from a cold guess to a warm one rarely runs in a straight line a person can picture. A player often knows the rough neighborhood long before finding the street, and the gap between those two states can swallow an afternoon.

The numbers bear this out. Players on the Hacker News thread report runs of well over a hundred guesses before a single word finally falls, with no penalty for the attempts beyond the time spent. Semantle treats that grind as the game itself rather than a failure state, which is either liberating or maddening depending on temperament.

The variants: Junior, Spanish, Teams, and copycats

The core game is punishing, so the site offers gentler doors. The FAQ points to Semantle Junior as an easier alternative, a Spanish edition served at the site's espanol path, and a Teams feature for playing with other people. Junior draws its secret words from a smaller pool of more common vocabulary, which trims the long tail of obscure answers that make the standard game so slippery.

Beyond the official variants, Semantle spawned a small ecosystem. The Hacker News thread highlights Pimantle, a reimplementation that plots guesses on a two-dimensional map so a player can see clusters and lines of similarity rather than reading a bare number. Others rebuilt the idea on different similarity models entirely. The concept travels well because the rule is so compact: hide a word, score by meaning.

The daily ritual and the archive

Part of Semantle's staying power is its rhythm. One word arrives each day and resets at a fixed hour, so the game is finite in time even though it is unbounded in guesses. The FAQ sets that reset at 8 PM New York time and points to an archive of past puzzles and an optional account, which let a committed player revisit old words or carry a record forward. None of this is loud. The site adds just enough structure to turn a hard puzzle into a habit, then gets out of the way.

Semantle also sits inside a wider wave of Wordle-descended daily games. The Wordle entry lists relatives such as Worldle and Quordle, and Semantle is often grouped with them as the semantic branch of the family. What unites them is the once-a-day cadence. What sets Semantle apart is that it refuses to tell a player how many letters they are even looking for.

How to actually play it without losing an afternoon

A field guide owes its readers a little practical advice, and Semantle rewards method over mashing. A common opening is to fire off a spread of unrelated common words, one from each broad category a person can think of: a concrete object, an emotion, a place, an action, an abstraction. The goal is not to guess the answer but to find which corner of the map is warm, then to walk toward it.

Once a guess breaks into the top thousand, the strategy flips from breadth to depth. A player should mine that region: synonyms, near-synonyms, and words that share the same contexts as the best guess so far. Rank, not raw similarity, is the compass here, because a jump from the high hundreds into the low numbers signals real progress in a way the percentage rarely does.

The last and hardest discipline is knowing when a rising number is lying. Because the model learned from news text, a cluster of related headline words can form a false summit. When several nearby guesses all plateau, the honest move is to retreat and probe a different sense of the word, rather than to keep climbing a slope that does not reach the peak.

Keeping notes helps more than it should. Because there is no board to read, the only memory of progress is the running list of guesses and their scores, and that list grows long. Experienced players effectively build a small map as they go, grouping their warmest words and testing the spaces between them, rather than trusting themselves to hold a 300-axis hunch in their head.

Who it is for, and why it endures

Semantle is not a game that produces a tidy shareable grid, and it was never trying to. It is for the kind of person who enjoys a slow, private puzzle, the same temperament that keeps a daily ritual going for its own sake. A site like Today's Rock, which presents a single rock each day and asks nothing more, draws on that same instinct: show up, pay attention to one small thing, come back tomorrow.

What has kept Semantle alive past the Wordle gold rush is that its difficulty is honest. The game is not hard because it withholds information through clever design. It is hard because meaning itself is tangled, and the model holds up a mirror to that tangle. Playing it is a short daily lesson in how machines have learned to measure language, delivered through nothing more elaborate than a text box and a number. For anyone curious about how modern language systems decide that two words are close, a few rounds of Semantle teach more than a diagram could.

FAQ

How does Semantle score a guess?

It converts both the guess and the secret word into 300-number vectors using the word2vec model, then measures the cosine similarity between them. The official FAQ adds that once a guess enters the thousand words nearest the answer, the site also shows its rank in a proximity column, which is far easier to read than the raw figure.

Is Semantle the same as Wordle?

No. Wordle, now owned by The New York Times, gives six tries at a five-letter word with color feedback on letters. Semantle hides a word of any length and scores purely on meaning, with unlimited guesses and no grid. It was created in 2022 as a harder counterpart, according to the Wikipedia entry on Semantle.

Why do some unrelated words score surprisingly high?

The vectors were trained on a large body of news text, so the model's sense of nearness reflects how words co-occur in journalism, not a dictionary definition. Words that often share headlines can land close together. The Hacker News discussion flags this as a recurring source of misleading signals.

Is there an easier version?

Yes. The FAQ points to Semantle Junior, which selects its secret words from a smaller set of more common vocabulary, along with a Spanish edition and a Teams mode for collaborative play. The main game remains the hard one by design.