The Best Wordle Strategy, Explained by Information Theory

Four green squares can feel like a near victory. Then you realize the answer could be FIGHT, LIGHT, MIGHT, NIGHT or RIGHT, and you have only two guesses left.

The problem is no longer finding common letters. It is choosing a move that separates the possibilities before your guesses run out.

Our recommendation is to start with a strong information-gathering word such as CRATE or TRACE, then adapt to the remaining answers and turns. Seek useful distinctions early, recognize groups of similar words, and switch to a direct answer when further investigation would waste a move. For players protecting a streak, the chance of finishing within six guesses should take priority over a small improvement in average score.

That recommendation draws on two explanations of Wordle through information theory: a 2026 paper by Talal Aladaileh and colleagues, and a science explainer by Sonia Choy for the Hong Kong Laureate Forum, written in Traditional Chinese. We also ran our own opening-word comparison using 1,880 distinct answers from our historical archive.

The result is a useful distinction: the best word for learning something immediately is not necessarily the best move for winning the whole game.

What makes a Wordle guess informative?

Every guess sorts the possible answers into groups. All answers in one group would produce exactly the same sequence of green, yellow and gray tiles.

Imagine eight equally likely answers remain. One guess divides them into two groups of four. Another divides them into four groups of two. Whichever feedback you receive, the second guess leaves fewer possibilities.

Information theory puts a number on that reduction. If a particular feedback pattern has probability p, the information it provides is:

I=-log₂ p.

A result that leaves four of eight equally likely answers gives one bit of information. A result that leaves two gives two bits. Under this equal-probability assumption, each bit corresponds to halving the candidate count.

This explains why an all-gray guess can still be valuable. If it eliminates six of those eight answers, it gives two bits, regardless of how disappointing the board looks. Tile colors tell you constraints; the number of green tiles does not, by itself, measure how much you have learned.

Entropy measures a guess before you play it

You cannot choose the feedback you receive. Instead, compare guesses by the information they provide on average, across every possible answer.

Let C be the current candidate set, with N equally likely answers. For a proposed guess g, let C_r contain the answers that would produce feedback pattern r. Then:

p_r=|C_r|N, H(g)=-∑r:p_r>0p_rlog₂ p_r.

This quantity, Shannon entropy, rewards guesses that distribute probability across many distinguishable outcomes, rather than concentrating it in a few large groups.

Five tiles have at most 3⁵=243 color patterns, although many patterns cannot occur for a particular guess and candidate set. A solver can calculate the probabilities by scoring the guess against every candidate and counting the resulting patterns.

If some answers are assigned higher probabilities, add their probabilities within each group instead of simply counting words. The entropy formula stays the same. What changes is the model of which answer you expect.

Two perspectives on the best Wordle starting word

The two sources share the same mathematical foundation, but they ask somewhat different questions.

The research paper: entropy versus a letter-frequency rule

Aladaileh, Donald Stephens, Mallak Alqaisi and Congyu Wu compare entropy-guided word selection with a baseline that favors letters appearing in many words. Their opening calculations use a 12,972-word dictionary, with TARES ranked first at 6.232 bits. Their top-ten table is worth a look: SOARE, the leader in our own archive comparison further down this page, appears fifth at 6.139 bits, and the eight words above and around it are all built from the same A, E, R, S and T. Their stated full simulation tests 2,315 solution words, using TARES as the first guess for both strategies. Read the paper.

The useful idea is to evaluate whole feedback patterns. Knowing that E is common helps, but it does not tell you whether a particular arrangement of letters will distinguish the answers left on your board.

However, the paper's numerical results need care. Section 5 reports a success rate above 99% for the entropy strategy and about 90% for the baseline, and those are the figures its press coverage repeated. Figures 13 and 14 of the same article instead label their win rates Win: 32.7% for the baseline and Win: 97.5% for entropy. These figures and the prose do not agree, so we cannot treat either pair as a settled estimate of the improvement.

Our reading refers to the open-access PDF as published at the DOI below, consulted on September 4, 2026. If a corrected version supersedes it, we will update this section and say so.

There is also a weak-baseline issue. Figure 12 shows the baseline playing LEAST on every guess from the second to the sixth against the secret word SLATE, while the entropy strategy solves it on the second. To the authors' credit, they identify this themselves: Section 6.1 notes that the letter-based strategy "can get caught up in an unending cycle" among words sharing the same five letters, naming LEAST, STALE, SLATE and STEAL exactly. That is a candid observation, but it also limits what the headline comparison can show. A method that replays a guess already known to be wrong is not just weaker than entropy; it is weaker than an ordinary attentive player, who would at least stop repeating a word the board has already ruled out.

We take the paper as an accessible motivation for information-based guessing, with unresolved reporting questions. It does not establish that its strategy is globally optimal, and it does not claim to: Section 6.2 states plainly that an optimal dynamic-programming approach solves any Wordle in at most five guesses, averaging about 3.42 with the best starting word, and that this "outperforms the greedy entropy method". The authors describe their own solver as a one-step greedy system without multi-round lookahead. That is the right reading of it, and it is the same distinction this article draws between a good immediate score and a good whole game.

The Hong Kong explainer: different goals produce different winners

Sonia Choy's article for the Hong Kong Laureate Forum, published in December 2023 and bylined to her role as a student editor of HKUST's science magazine Science Focus (《科言》), makes the multi-turn distinction especially clear. It is written in Traditional Chinese. It reports different leaders for different objectives:

These are historical, source-reported results under the word lists and strategies described in the Hong Kong Laureate Forum article. We have not independently reproduced those full-game averages.

They did not originate there. The SOARE and SALET results trace back to Grant Sanderson's 3Blue1Brown videos on Wordle and to Alex Selby's exhaustive search, which reported SALET as the optimal opener for the original 2,315-word answer list at an average of about 3.421 guesses. The 3.412 quoted above is slightly lower than Selby's figure, which is a further reason to treat any single decimal place here as specific to one particular setup rather than as a fixed property of the game.

The difference between 3.412 and 3.434 is 0.022 guesses per game: roughly one extra guess across 45 games, on average, under those simulations. That helps put the search for a perfect opener into perspective. A familiar word can be a sensible practical choice even when a specialized analysis gives a slightly better score to an obscure one.

Why TARES, SOARE and SALET can all lead a ranking

Before comparing rankings, ask three questions:

1. Which words can be answers? A historical solution list, a broad dictionary and a current game's eligible answer pool are different populations.
2. Which words can be guessed? An information-gathering guess may be useful even if it cannot be the answer. Hard Mode also restricts later choices.
3. What is being optimized? Immediate information, average candidates left, expected total guesses and the chance of finishing within six turns are different objectives.

The paper's 12,972-word probability model and Choy's discussion of 2,315 possible answers therefore need not produce the same first-choice word. Nor should their entropy values be compared as if they came from one benchmark.

It also matters how probabilities are estimated. Choy introduces the idea with general English letter frequencies and explicitly notes that a Wordle-specific calculation would be more accurate. The simplification is useful for intuition, but letter frequency in running text is not the probability that a five-letter answer contains that letter. Letters within words are not independent, either.

For exact calculations, simulate complete feedback against complete words. Multiplying generic letter-absence percentages cannot fully capture positions, spelling patterns or repeated letters.

Repeated letters need particular care. If one E is green and another E is gray, that gray tile can rule out an additional E rather than rule out E altogether. Correct scoring allocates green matches first, then yellow matches from the remaining copies of each letter.

Our calculation: 11 opening words, 1,880 historical answers

To make the ideas concrete, we compared 11 selected openers against 1,880 distinct historical answers in our site's archive snapshot, including answers recorded through September 3, 2026.

Each answer received equal weight and appeared once. We calculated exact feedback, including repeated letters. We retained every answer in the snapshot, without filtering out words absent from a general English frequency dataset.

This is a retrospective comparison on a fixed vocabulary. It is not a simulation of today's eligible answer pool, a search over every allowed opening word, or a measurement of our users' scores.

Here is the full result. Every figure below is produced by the script linked at the end of this article, run against the frozen snapshot.

Source: WordleSolver Pro's original calculation on the frozen archive snapshot. “Candidates left” counts the compatible answer itself, including a one-word group when the guess is correct. The method is described at the end of this article.

See it for yourself

The table records the answer; the experiment below shows the working. Choose an opener and select Watch it work to see all 1,880 answers sort themselves into feedback groups, then open any group to read the words still standing. Compare all 11 recomputes the whole table on your device, and the endgame panel lets you switch between FIGHT and FLAME to see why their chances differ.

Open the experiment on its own page if you would rather explore it full width, or share it on its own.

Four findings matter for ordinary play.

First, familiar openers perform well. SOARE has the highest entropy among these 11 words, but TRACE and CRATE are close behind. The case for CRATE as a practical default is not that any single analysis crowns it: the full-game comparison Choy reports is actually led by SALET. It is that CRATE is never far from the top under either measure. It gains more information than SALET on our archive (5.873 against 5.795), and it gives up about 0.022 guesses per game to SALET in that full-game comparison. A word you can remember, that costs you roughly one extra guess every 45 games, is a good default. TRACE is an equally defensible choice. Neither result proves a universally best opener.

Second, the paper's top-ranked word finishes last among the strong openers here. TARES leads the paper's 12,972-word ranking, yet it places eighth of the eight balanced openers in our archive comparison, at 5.734 bits against SOARE's 5.902. Nothing is wrong with either calculation. They are scoring the same word against different populations, which is exactly the point of the previous section: an opening-word ranking is a statement about a word list, not about Wordle.

Third, testing four vowels comes at a cost. ADIEU and AUDIO leave approximately 98.5 and 147.1 candidates on average, compared with CRATE's 55.4. Testing useful consonants as well as vowels can distinguish more answers. These are differences in immediate narrowing, not measured differences in eventual win rate.

Fourth, maximizing entropy is not the same as minimizing the average candidate count. TRACE has slightly higher entropy than IRATE, yet IRATE leaves fewer candidates on average. Maximizing entropy minimizes the average logarithm of the remaining group size under this uniform model. The candidate-count column averages the sizes themselves, giving large groups more influence.

This also explains why counting green tiles, avoiding all-gray rows or testing the most new letters cannot serve as a complete strategy. Each is only a partial description of what a move accomplishes.

The endgame: when an “impossible” answer is a useful guess

Return to FIGHT, LIGHT, MIGHT, NIGHT and RIGHT. Suppose these are the only remaining answers and each is equally likely.

Guessing FIGHT either wins immediately or leaves four indistinguishable alternatives. If you have only two turns left, a second candidate guess gives you a total success probability of 2/5=40%.

In normal mode, consider FLAME instead. It cannot solve the puzzle immediately, but its feedback separates the candidates:

With two turns left, FLAME lets you solve three cases with certainty on your final guess and choose between the remaining pair. The overall success probability is 4/5=80%.

With three turns left, FLAME guarantees a finish in this five-answer example: one probe, followed by at most two candidate guesses. These percentages are exact calculations for this specified situation, not player statistics.

The lesson is to test the letters that differ between candidates. Reconfirming the shared ending provides little help.

Why the same idea stops working with one or two candidates

If only NIGHT and RIGHT remain, simply guess one. Under equal probabilities, that takes 1.5 further guesses on average: one guess half the time, two guesses the other half. A non-answer probe that identifies the correct word would still require a second move to enter it.

With only one turn left, the rule is even simpler: guess the most probable remaining answer. A word that cannot be the answer has zero chance of winning, however informative its feedback would be.

Information helps because it improves later decisions. When there is no later decision available, information alone has no value to your score.

Hard Mode changes the available strategy

Hard Mode requires subsequent guesses to use revealed hints. If the ending IGHT is already green, FLAME is unavailable because it abandons those positions.

The normal-mode calculation above therefore does not transfer to that Hard Mode state. The restrictions themselves can make a family of near-identical answers dangerous. Planning for those branches earlier matters more when you cannot freely play a diagnostic word later.

The strategy we recommend

For everyday players, our preferred approach is adaptive play with the six-guess limit in mind:

1. Open with a strong, familiar word. CRATE is our practical default; TRACE is a close alternative in our archive calculation. SLATE and RAISE are also reasonable choices. Spend more attention on interpreting the next board than on tiny opening-score differences.
2. Apply every clue before choosing again. Track green positions, yellow position exclusions, gray letters and letter counts. A contradictory candidate should leave the list.
3. Look at what separates the remaining words. If many share an ending or several fixed positions, test their differences while enough turns remain. In normal mode, this can justify a guess that is not a possible answer.
4. Use the remaining turns to decide when to solve. With two equally likely candidates, guess one. On the final turn, choose the most likely answer. Do not spend a move gathering information you cannot use.
5. Respect the mode you are playing. A normal-mode recommendation may be illegal in Hard Mode, and the best continuation can change with that restriction.

For a computer solver, the corresponding goal is to evaluate the whole decision tree: possible feedback, subsequent guesses and whether each branch finishes in time. Greedy entropy is a useful guide, but it evaluates only the immediate information return.

A claim of mathematical optimality needs a specified answer set, answer probabilities, legal guesses, game mode and scoring objective. For streak protection, we would prioritize the probability of finishing within six guesses, then use expected guess cost to distinguish equally reliable strategies. For someone chasing a three-guess average and accepting occasional losses, the preferred trade-off may differ.

There is also an implementation detail that matters: information scores must reflect the current candidates. With two equally likely answers left, their uncertainty is only one bit. No next guess can reveal more than that. Figure 10 of the paper reaches a state it labels "Next word options: 2" and lists PLUMB at 3.932 bits and FLUMP at 3.766 beside it. No guess can be worth 3.9 bits when only one bit of uncertainty remains, so those displayed values cannot be conditional entropies of that two-word state; they look like scores carried over from the full word list. Whatever scoring convention produced them, a solver should distinguish a word's initial score from its value after the latest feedback.

You can use our Wordle solver to enter the board's clues and inspect compatible words. Its candidate filtering and frequency ordering are useful aids; the opening-word experiment in this article is a separate calculation, not a claim that the live result list implements an optimal decision tree.

What our site data can tell us

Our Wordle statistics page uses a lexical difficulty model and quality-checked search-demand observations. That offers a way to study which puzzles attract help-seeking and generate questions for further investigation.

Search demand does not reveal a player's opening word, complete sequence of guesses or final result. It can also change with our visibility in search results. Even complete records from solver visitors would describe a self-selected group rather than a random sample of all Wordle players.

We therefore do not use those scores to claim that a particular opener increases player win rates. The original numerical evidence here comes from our historical answer vocabulary and exact hypothetical endgames.

A stronger real-player comparison would need opening guesses, subsequent moves, game mode and outcomes, including failures. It would also need to account for puzzle difficulty and player experience. Otherwise, differences between players could be mistaken for differences between strategies.

Frequently asked questions

What is the best Wordle starting word?

CRATE is our practical recommendation, with TRACE a close alternative. SOARE has the highest first-guess entropy among the 11 words in our archive comparison, while the analysis Choy describes reports a lower full-game average for SALET. These results answer different questions and do not establish one winner for every word list and game mode.

Is ADIEU a bad first guess?

It can provide useful clues, but it is less effective at narrowing our historical answer set than several balanced vowel-and-consonant openers. ADIEU leaves 98.5 compatible answers on average in our calculation, compared with 55.4 for CRATE. That measures the first guess, not a player's overall skill.

Should I always use the same first two words?

A fixed pair is easy to remember, but it can spend the second turn testing letters that the first result has made irrelevant. An adaptive strategy can still choose that same second word when useful and choose a better one when the feedback calls for it. Reassess after the first guess.

Should I avoid repeated letters?

Five distinct letters often make an efficient opener, but repeated letters are sometimes essential to distinguish later candidates. Let the remaining answers determine whether testing another copy of a letter is useful.

Does the highest-entropy guess guarantee a win?

No. It maximizes expected immediate information under the chosen probability model. It does not automatically minimize total guesses or guarantee that every feedback branch can be solved before the turn limit.

Sources and reproducibility

• Aladaileh, T., Stephens, D., Alqaisi, M., and Wu, C. (2026). Solving Wordle Using Information Theory. Northeast Journal of Complex Systems, 8(1), Article 6. Relevant locations: the word counts in Section 3; the opening ranking in Figure 8; the worked game in Figure 10; the strategy comparison in Figure 12; the reported success rates in Section 5 and in Figures 13 and 14; the authors' own observations in Sections 6.1 and 6.2. The conflict between the reported success rates is discussed above.
• Choy, Sonia 蔡蒨珩 (December 2023). Demystifying Wordle: A Crash Course in Information Theory. Hong Kong Laureate Forum, in Traditional Chinese; bylined to the author's role as a student editor of Science Focus (《科言》), the science magazine of HKUST's School of Science. The article reports the SOARE, SLANE, SALET and CRATE comparisons quoted here; those results originate with the analyses described in the next entry.
• Sanderson, Grant (3Blue1Brown). Solving Wordle using information theory. The treatment that put entropy-based Wordle analysis in front of a general audience, and the origin of the SOARE and SALET results quoted second-hand above.
• Selby, Alex (2022). The best strategies for Wordle. An exhaustive search over the original 2,315-word answer list, reporting SALET as the optimal opener at an average of about 3.421 guesses. This is the underlying source for the full-game figures, and its scope is a useful reminder that "optimal" is defined relative to one specific word list.
• WordleSolver Pro archive experiment (September 4, 2026). Our own calculation, run against a frozen snapshot of the distinct answers in our past Wordle answers archive through September 3, 2026.

The archive experiment uses 1,880 distinct answers and 11 specified guesses, with a uniform probability of 1/1880 per answer. Repeated historical appearances do not receive extra weight. All 20,680 guess-answer pairs are scored with two independently written implementations of duplicate-aware feedback, and every result is checked for agreement between them before publication.

For a feedback group of size n_r, its probability is n_r/N. The table's average remaining count is ∑_r n_r²/N; its largest group is max_r n_r. These are exact enumerations for this finite snapshot, rounded for display. Sampling confidence intervals would not address the central limitation: a different answer population or probability model can produce different rankings. We did not run a full-game benchmark or prove optimality over all legal guesses.

The interactive experiment on this page uses that same checked calculation and the same frozen answer set, so the figures it produces match the table above exactly. It runs in your browser, with nothing to install.