Wordle Solver — Best Next Guess by Information Gain

Enter your guesses and the colours the game gave back. It filters the candidates and ranks what to guess next by expected information gain — which word teaches you the most — rather than only listing what is still possible.

Runs entirely in your browser. Nothing is uploaded, stored, or sent anywhere.

How to use it

Type the word you guessed into the box, then click each of the five tiles to cycle it through grey, yellow and green until they match what the game showed you. The results update as you type. Add a row per guess. Nothing is sent anywhere — the whole five-letter lexicon is embedded in the page, so it works offline once loaded.

What “expected bits” means

The useful question is not “which word might be the answer” but “which guess will teach me the most.” Every guess splits the remaining candidates into groups, one per feedback pattern it could come back with. A guess that splits 300 candidates into many small groups tells you far more than one that leaves a single group of 250 — even if the second word is more likely to be right.

The measure for that is Shannon entropy. If a guess returns pattern i with probability p_i, its expected information is

One bit means the candidate list is expected to halve. A guess worth 6 bits cuts the field by roughly a factor of 64. The average left after column is the same calculation read the other way round — the expected number of candidates once the feedback arrives — which many people find easier to interpret than bits.

The last column matters when you are near the end. A guess that is itself a possible answer can win outright this turn; one that is not can only ever gather information. When two guesses are close on bits, the one that could be the answer is usually the better play, which is how the ranking breaks ties.

Openers

With no guesses entered, the table ranks opening words. Those figures are computed against every one of the 8,636 five-letter words in the lexicon — 74.6 million guess/answer evaluations — which is why they are precomputed rather than worked out in your browser.

tares comes out on top at 6.22 bits, ahead of a tight cluster of near-anagrams: lares, rales, rates, tales. The pattern is not an accident. They pair the highest-frequency consonants with two common vowels and a trailing s, so they cut the field more evenly than a word carrying a rare letter. The spread across the top ten is under 0.25 bits, so anything in that cluster is a perfectly good opening and the difference over a full game is well below one guess.

Repeated letters, where most solvers get it wrong

The feedback rule for a repeated letter is subtle, and it is the single most common bug in home-made solvers. Green tiles are assigned first, and each one consumes a copy of that letter in the answer. Only then are the remaining positions scanned left to right, and each takes a yellow only while an unconsumed copy is still available — otherwise it comes back grey.

So guessing geese against steel returns grey, yellow, green, yellow, grey. The final e is grey even though the answer plainly contains an e, because both of the answer's copies have already been spoken for. This implementation follows that rule exactly and is checked against hand-worked cases — including speed/erase, abbey/kebab and array/radar — before the page is built.

Which word list this uses

The candidate set is the 8,636 five-letter words in ENABLE, a public-domain lexicon. The game's own answer list is a much smaller, hand-curated subset of common words and is not reproduced here. The practical consequence is that this tool considers obscure words the game would never actually use, so its candidate counts are conservative: the real field is smaller than the number shown, and the true answer is usually one of the more ordinary-looking words in the list.

Notes

• When the candidate list is still large, the tool ranks next guesses using the candidates themselves rather than the whole lexicon, to keep the page responsive. It states which pool it searched underneath the table rather than narrowing silently.
• “No word matches that feedback” almost always means a tile colour is set wrong. Repeated letters are the usual culprit.
• Wordle is a trademark of The New York Times Company. This tool is not affiliated with or endorsed by them; the name is used only to identify which puzzle the feedback format belongs to.

More free tools

Each of these runs entirely in your browser. Nothing is uploaded, stored or sent anywhere, and none of them needs an email address.

See all 49 engineering tools →