The idea
A guess splits the set of still-possible answers into buckets, one per colour pattern. Shannon entropy measures how evenly it splits them:
H(guess) = − Σ p(pattern) · log₂ p(pattern)
A guess worth 5.9 bits cuts the candidate pool by a factor of about 25.9 ≈ 60 on average. The solver scores every word in the dictionary this way — including words that cannot be the answer — and plays the highest scorer, breaking ties in favour of words that could still win outright.
What this build does
Two candidate pools. The curated answer list (2,309 words, what the real game draws from) starts you at log₂2309 = 11.17 bits of uncertainty. The full legal-guess list (12,953 words, including aalii and friends) starts at 13.66 bits and is a much harder game. Guesses may come from the full 12,953 either way.
Turn-one rankings are precomputed for both pools — they never change — so the first move is instant; later turns are scored live, in slices, so the page stays responsive on a phone.
Measured performance
Curated pool, opening with soare: all 2,309 answers solved within six guesses, average 3.53. Distribution — 2: 31, 3: 1103, 4: 1107, 5: 67, 6: 1.
Full legal pool, opening with tares: a 250-word random sample solved 100% within six, average 4.10. Distribution — 2: 1, 3: 43, 4: 137, 5: 67, 6: 2. Five times the haystack costs barely half a guess, which is what a logarithmic measure of uncertainty predicts.
Source
T. Aladaileh, D. Stephens, M. Alqaisi, C. Wu, “Solving Wordle Using Information Theory,” Northeast Journal of Complex Systems 8(1), 2026. doi:10.63562/2577-8439.1146 — reported 99% solve rate versus 90% for a common-letter strategy.
Using it on the real game
Open Solve a real game, type the word you actually played, tap its tiles until the colours match the ones Wordle showed you, then press Add clue. The next recommendation appears immediately.