What is the best Wordle starting word?
Last verified 08-07-2026
SALET. Played against all 2,380 possible Wordle answers, it solves 100% of them in an average of 3.4403 guesses.
That is every answer, with none removed for having been used before. The New York Times has been reusing answers since 02-02-2026, and a solver that drops used words fails on exactly those days.
The more interesting result: the popular method picks the wrong word
The usual way to choose an opener is to maximise expected information — how many bits the first guess reveals, on average, about which answer it is. By that measure the best opener is SOARE, at 5.889 bits.
SOARE is not the best opener. Measured by the thing that actually matters — how many guesses it takes to finish — it comes fifth.
| Opener | First-move bits | Rank by bits | Average guesses |
|---|---|---|---|
| SALET | 5.836 | 7 | 3.4403 ← best |
| SLATE | 5.855 | 6 | 3.4429 |
| TRACE | 5.817 | 11 | 3.4450 |
| CRANE | 5.727 | 36 | 3.4466 |
| SOARE | 5.889 | 1 ← best | 3.4710 |
| RAISE | 5.882 | 2 | 3.4807 |
The two columns disagree. SALET ranks seventh on information and first on results; SOARE ranks first on information and fifth on results.
Why entropy misleads here
Expected information is a greedy measure: it scores a guess by what it reveals right now, with no regard for what happens afterwards. A word can split the field beautifully on turn one and still leave you in positions that are awkward to finish from — clusters of words differing by a single letter, where each subsequent guess can only eliminate one candidate at a time.
SALET gives up about 0.05 bits on turn one and lands in better endgames. Over 2,380 puzzles that is worth roughly 0.03 guesses, which is the difference between first and last place.
The general lesson: optimise the metric you care about, not the one that is easy to compute.
What happens after the opener
Given a fixed opener there are only 243 possible feedback patterns, and only 149 of them can actually occur after SALET. That means the entire second move can be solved in advance and stored in a 7KB table — so the first two turns are instant, and by the third the field has usually collapsed to a handful of words.
The solver uses exactly this, and shows the bits each suggestion gains so you can see why it recommends what it does.
Distribution
Across all 2,380 answers, with SALET:
| Guesses | Answers |
|---|---|
| 2 | 80 |
| 3 | 1,224 |
| 4 | 1,017 |
| 5 | 47 |
| 6 | 3 |
| Failed | 0 |
How this was measured
The answer pool is the union of the leaked original curated list and every answer the New York Times has actually used — 2,380 words, because NYT has used 65 words the original list never contained. The guess pool is 12,972 words.
Every figure here comes from playing all 2,380 possible answers to completion with the solver on this site. Nothing is estimated and nothing is quoted from another site. You can re-derive it: the repository is public and npm run test:solver prints the solve rate and average.
Related research
- Does Wordle repeat answers? — yes, since 02-02-2026, and it breaks most solvers.
- How does the NYT Wordle API work? — it serves the full history and about a month ahead.