Five-letter word solver

Repeated letters: which copy comes back yellow

One thing about this game is misread more often than everything else put together, and it is what happens when your guess uses the same letter twice. The game does not mark letters; it marks tiles, and it does it in two passes. First, every tile whose letter is already in the right place turns green — and that copy of the letter is used up in the answer. Then the tiles that are left are read from left to right, and each one turns yellow only while an unused copy of its letter is still there; once there is none left, the tile is grey.

Work one all the way through. You guess allot and the answer is banal. No tile matches in place, so nothing turns green and all five letters of the answer are still unused. Now left to right: the a of allot claims one of the answer’s two a’s and turns yellow; the first l claims the answer’s only l and turns yellow; the second l finds no l left and turns grey; o and t are nowhere in banal, so both are grey. Yellow, yellow, grey, grey, grey — and the two l’s came back differently although they are the same letter in the same guess.

That is what grey means on a repeated letter, and it is not what most people read it as. It does not say the letter is missing from the answer. It says there are no more copies of it than the ones already marked yellow or green in that row. Which copy gets the yellow and which the grey is decided entirely by that left-to-right walk, running against whatever the greens have not already spent — so the same guess can hand you a yellow before a grey or a grey before a yellow, depending on where the green sits.

• With no green anywhere, the earlier copy is the one that scores. Guess allot against banal and the first l is yellow while the second is grey, because the walk spends the answer’s only l on the copy it reaches first.
• A green further along the row takes its copy before the walk even begins. Guess sassy against visit and the s in third position is green while the s in first position is grey — the green is settled in the first pass, so nothing is left for the earlier tile. Guess eerie against heavy and the same thing happens on both sides at once: the e in second position is green, and the e before it is grey along with the one after it.
• A green earlier in the row does exactly that to the copies behind it. Guess sassy against sight and the first s is green while both later copies are grey — the reverse of the first case, out of the same rule rather than out of a different one.

None of this means only one copy ever counts, which is the second misreading. Guess geese against sheep: the e in second position is yellow, the one in third is green and the last is grey — three different marks for one letter in one row, plus a yellow s. Guess abbey against hobby: the first b is yellow and the second is green, because the answer holds two b’s and both of your copies are accounted for. Count the yellows and greens for a letter and you have counted how many of it the answer holds; that is the fact the tool works from.

The order is also why this tool sometimes refuses a board instead of answering it. Within one row a repeated letter can never be grey before it is yellow, because the walk spends what it finds as it goes — the yellow copies always come first. A board that says otherwise describes no word at all, and it is nearly always this section’s own example with two tiles swapped, so the tool names the two tiles rather than handing back an empty list. That check is load-bearing rather than tidy: the tool does not re-mark all 8,636 words against your board. It compiles the board once into what a word must contain and where, and asks each word about that. The compiled question and the marking rule are the same question only because grey-before-yellow has been ruled out first; without it, the compiled version would admit words the game could never have marked that way.

Why our word list is not the game’s

The words here are ENABLE’s: the 8,636 of exactly five letters, out of a public-domain list of 172,820 that a great many word games and puzzle sites are built on. It is not the list the puzzle itself uses, and it was never going to be. The game’s own accepted-guess list and its much smaller answer list belong to its publisher, and the answer list is curated and edited over time, so even a copy of it would agree only for a while. What you get instead is a list anybody can check, and a tool that says out loud where it and your game part company.

• It will sometimes suggest a word your game refuses. That is friction rather than damage: the game will not submit a word outside its own list at all, so no guess is spent — you type a different word and carry on.
• It will sometimes not know a word your game accepted. Enter that row anyway and mark it exactly as the game did: the marks came from the game, so the row narrows the list as far as any other row does, and our not holding the word costs you nothing at all here. The list decides which words can be suggested back to you, and nothing about which of your rows are read.
• It holds words that will never be an answer. ENABLE is a dictionary rather than an answer list, so plurals, verb forms and thoroughly obscure entries all sit in it — which is why a suggestion can be perfectly consistent with your board and still be a word you would not choose to type.

So read what you find here the way you would read a dictionary propped at the side of the table: a strong shortlist, checked against the game you are actually playing. The disagreement runs in both directions and neither side of it is a defect to be fixed — it is what happens when the reference you may legally use is not the reference the puzzle uses. And to say the obvious plainly: this tool is not affiliated with, endorsed by or connected to the publisher of any word puzzle. It works from the game’s rules, which belong to nobody, and from a word list of its own.

The number beside each word, and why the list is not sorted by it

Every row carries one number: how many words would still be possible afterwards, on average, if you guessed that word next. It is worked out by marking that word against every word still standing, grouping them by the pattern of marks that would come back, and averaging the size of the group you would land in. A row reading 18.4 is telling you that whatever the game answers, you should expect about eighteen words left — and the last row of a finished list reads 1.0, which is exactly right: guess it and one word remains, and that word is the answer.

The list is not sorted by that number, and the page says so above the list rather than leaving it to this page. The order is by how much a guess would teach, which counts many small groups as worth more than a few large ones — two honest readings of the same split that disagree at the margins. So the column wobbles instead of running in order, and the opening words are the plainest case: the first four are tares at 195.9, lares at 185.6, rales at 187.6 and rates at 201.0.

Why show the average at all, then, instead of the quantity the order is actually made of? Because that quantity is measured in bits, and a column of bits is a course in information theory standing between you and your next guess. "About eighteen words left" is a sentence you can act on. Use the order to choose and the number to judge: a large number early on means no single guess can do much yet, and a number that drops sharply means the top of the list is doing real work.

Opening words, and why the best one is not a matter of taste

An empty board tells the tool nothing, which makes it the one question with the same answer for everybody — so it is answered once, when the site is built, rather than in your browser. Every one of the 8,636 words is marked against the whole list, itself included; the splits are counted, and the fifty best are shipped inside the page in all thirteen languages. That is why they are on screen before you have typed anything and why they cost you no download at all.

• They are the best openings for this list, not for the game’s. Words you have seen recommended elsewhere — crane, slate, adieu — are all in here and all ranked, but they are ranked against our 8,636 rather than against the few thousand words the puzzle draws its answers from.
• The top few are close to interchangeable. The distance between the first opening and the tenth is small next to the difference your second guess will make, so take one you can remember and spend the attention on the move after it.
• What makes them good is not that they are familiar words but that their letters are. The strongest openings are built from a, e, r, s and one of l or t, spread over positions where they discriminate — which is why so many of them, tares and rales and lares and reals among them, are anagrams of one another.

There is one stronger play this tool deliberately does not make. Sometimes the most informative guess is a word that cannot possibly be the answer — a word chosen only to separate the candidates that are left. Every suggestion here is a word that could still be the answer, which costs a little in theory and buys two things in practice: the tool never spends your guess on a word that cannot win, and everything it suggests is legal in Hard Mode.

Hard Mode: what changes, and what does not

Hard Mode makes the game refuse any guess that does not reuse what it has already shown you: every green letter must stay where it is, and every yellow letter must appear somewhere in the word. It restricts what you are allowed to type; it changes nothing whatever about how a guess is marked, so the repeated-letter rule above holds exactly as it stands.

• Every word this tool lists is already a legal Hard Mode guess. It only ever shows words that could still be the answer, and such a word keeps every green in place and carries every yellow letter — so you can take the top of the list without checking it against the rule.
• What Hard Mode takes away is the guess that cannot be the answer. In ordinary play, with a board narrowed to batch, catch, hatch, latch, match, natch, patch, ratch and watch, you can spend a turn on a word carrying four of those first letters at once and be certain afterwards. Hard Mode forbids it, and the nine words have to be walked one at a time.
• That is also what the numbers are telling you there. Every one of those nine leaves 7.2 words on average, because none of them separates the others — when the list is short and the words rhyme, an identical column across every row means there is nothing clever left to do, in Hard Mode or out of it.

The tool does not filter for Hard Mode, and it does not need to: the whole list is legal there already. What it cannot do is warn you that a board is about to become the trap above, so that is the judgement to keep for yourself. When several rhyming words are still standing and you have guesses to spare, an ordinary game lets you break them apart in one move; in Hard Mode the only thing to do is start guessing, and the order on this page is the best order to guess them in.

How we worked out the best starting word

The best Wordle starting word is a question with an arithmetic answer, and the answer on this page was computed here rather than taken from anybody. What we did, in order: take ENABLE’s public-domain word list, keep the 8,636 entries of exactly five letters, and then, for each of those words in turn, mark it against all 8,636 by the two-pass rule at the top of this page, group the outcomes by the pattern of marks that would come back, and score the split that results. Do that 8,636 times over and the words sort themselves. The one that comes out first is tares.

• The corpus is the whole five-letter part of that list, with nothing taken out of it. There is no shortlist of likely answers, no cut by how common a word is and no hand-picking: every five-letter entry ENABLE holds is at once a guess that gets scored and a word the other guesses have to separate.
• Every word is treated as exactly as likely to be the answer as every other. That is the assumption the average rests on, and it is a choice rather than a fact about the puzzle — but it is the choice that needs nothing about the game beyond its marking rule, which is what makes the whole calculation something anybody can repeat.
• Two words fall in the same group when the guess would come back marked identically against both, so a group is a set of words that one guess cannot tell apart. What a guess is worth is what it does to the size of the group you would land in.
• Ties are settled by the word itself, and the group sizes are added up smallest first, so the same list of words always produces the same ranking down to the last digit. That reads like housekeeping and is not: it is what makes it possible for a second program to check the first.

Nothing here is quoted from anybody else’s study, which means these numbers can be wrong and can be caught being wrong. Every one of them is recomputed from the shipped word list by a second program that shares no line of code with the one this page uses — a different way of marking a guess, a different way of counting the groups — and the two have to produce the same fifty words with the same averages, to the last digit, before any of it ships. Every pair of words used as an example anywhere in this guide was run through that same marker, for the same reason. The one thing that cannot be checked here is the calculation against the puzzle’s own answers, and the next section is about what that costs.

What the ranking settles, and what it does not

What it settles is where the words other people recommend actually land against this list. Take slate: it is among our fifty and 34th in the order, and the number beside it reads 247.3 where tares reads 195.9. The two other openings you will have seen suggested, crane and adieu, are not among the fifty at all. None of that is a verdict on them. A word is only ever good against some list of words, and which list that is is the first thing this ranking cannot settle for anybody else. But it is a measurement rather than an opinion, and it is what this list says.

• It measures one guess and not a game. The number is how many words would still be standing after that opening, and nothing on this page plays out the rest of the puzzle or claims an average number of turns. A strong opening buys a smaller list to work from; the guess you make against whatever comes back is where most of the game is decided.
• Treating every word as equally likely has a direction, and it shows in the answer. Our list is a dictionary rather than an answer list, and a dictionary is thick with plurals, so a word carrying an s cuts it well — every single one of the fifty openings carries one. That is the corpus showing through rather than a fact about the puzzle, and it is one honest reason why our openings and the ones tuned against a curated answer list do not agree.
• It is settled once and learns nothing afterwards. The ranking is never adjusted from what yesterday’s answer turned out to be, and the tool never sees an answer at all, so it cannot get better at one particular puzzle’s taste in words — and cannot quietly get worse at it either.

That is the trade a page computing its own numbers makes. A figure quoted from somebody else’s study may well have been taken against the puzzle’s real answers, which is a thing this page cannot do; what you get here instead is a number you can recompute from a word list anybody may download, and a page that is plain about giving a dictionary’s answer rather than the puzzle’s. Take the openings as the strongest first cut available without touching anybody’s property, and take the section on the word list above for where the two part company.