How the Nerdle Solver Works

When you play a guess, every possible answer would produce some specific coloring — greens, purples, blacks. That means a guess partitions the suspect list: every answer falls into exactly one feedback-pattern bucket, and once the real colors appear, every answer outside your bucket is eliminated instantly.

So a guess is exactly as good as the buckets it creates. Many small buckets mean that whatever feedback you get, few suspects survive. A few big buckets mean you can get unlucky and learn almost nothing. Our first-ranked model opener, 52-34=18, splits the field into 1,703 buckets. A repeat-heavy guess like 11+11=22 manages only 415 — it asks the same question about the same few characters twice.

To build its candidate pool, the solver first asks: if the answer is chosen at random, how many suspects will still be alive after I see the feedback? Big buckets get weighted by how likely you are to land in them (a bucket holding 100 answers is hit 100 times as often as one holding 1), which is why lopsided guesses score badly.

The live solver then calculates expected total guesses for the visible candidates and sorts the first suggestion page by that full-game metric. Expected answers left remains the exact one-move comparison shown beneath each suggestion.

Same rules, same 17,723 suspects — but the first-ranked opener leaves you a to-do list of ~33 while the worst of these leaves ~319. The first guess changes how much work remains, but the later choices matter too. Compare the results in the best starting equations.

Scoring one guess is cheap; searching every legal strategy at every turn is not. We therefore build a practical strategy offline and save it as a decision tree whose root is 52-34=18 — the recommended opener, currently ranked #1 by exact expected total guesses in our opener ranking. The tree has one branch per feedback pattern, each leading to the strategy’s next reply, and so on until every branch ends in the answer.

At each node, the builder pools every remaining answer as a candidate guess, and once 500 or fewer answers remain it adds up to 2,000 deterministically sampled probe guesses. The 64 candidates that leave the fewest expected answers after one move then receive an exact expected-total-guesses evaluation, and the winner is chosen with the same tie-break chain the live suggestion ranking uses. That shortlist keeps the tree tractable, but it also means we do not claim a mathematical global optimum. The full methodology records the dataset, ranking rules, and limitations in one place.

The published opener table goes one step further: it treats every modeled answer form as a possible opener, plays a fast greedy follow-up strategy against every modeled answer, and measures complete game length. Those averages are exact for that follow-up strategy — the committed tree’s exact per-node evaluation averages slightly fewer guesses. The tree sometimes plays probe guesses that can’t be the answer (they appear throughout the best-second-guess tables). When you use the solver, a position on the known path can walk this precomputed tree. If your guesses leave that path, the solver ranks candidates for the remaining pool instead.