Original research · FiveLetterWords · through 2026-11-05
Prepared . This snapshot is frozen.
Where Wordle answers come from: data release
The completed archive contains 1966 puzzles, 1937 different answers and 29 repeats. Among the latest 90 answers, 30 are outside the frozen September 2026 solver pool. The unused source categories now contain 440 original-list words and 785 July 2023 additions.
The monthly mix
Monthly counts and source categories
| Month | Puzzles | U_orig | U_2023 | Used_2021 | Used_later | Other |
|---|---|---|---|---|---|---|
| 2023-01 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2023-02 | 28 | 28 | 0 | 0 | 0 | 0 |
| 2023-03 | 31 | 30 | 1 | 0 | 0 | 0 |
| 2023-04 | 30 | 29 | 1 | 0 | 0 | 0 |
| 2023-05 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2023-06 | 30 | 28 | 2 | 0 | 0 | 0 |
| 2023-07 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2023-08 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2023-09 | 30 | 30 | 0 | 0 | 0 | 0 |
| 2023-10 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2023-11 | 30 | 30 | 0 | 0 | 0 | 0 |
| 2023-12 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2024-01 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2024-02 | 29 | 29 | 0 | 0 | 0 | 0 |
| 2024-03 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2024-04 | 30 | 29 | 1 | 0 | 0 | 0 |
| 2024-05 | 31 | 30 | 1 | 0 | 0 | 0 |
| 2024-06 | 30 | 30 | 0 | 0 | 0 | 0 |
| 2024-07 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2024-08 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2024-09 | 30 | 29 | 1 | 0 | 0 | 0 |
| 2024-10 | 31 | 30 | 1 | 0 | 0 | 0 |
| 2024-11 | 30 | 28 | 2 | 0 | 0 | 0 |
| 2024-12 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2025-01 | 31 | 29 | 2 | 0 | 0 | 0 |
| 2025-02 | 28 | 27 | 1 | 0 | 0 | 0 |
| 2025-03 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2025-04 | 30 | 29 | 1 | 0 | 0 | 0 |
| 2025-05 | 31 | 29 | 2 | 0 | 0 | 0 |
| 2025-06 | 30 | 29 | 1 | 0 | 0 | 0 |
| 2025-07 | 31 | 26 | 4 | 0 | 0 | 1 |
| 2025-08 | 31 | 28 | 2 | 0 | 0 | 1 |
| 2025-09 | 30 | 30 | 0 | 0 | 0 | 0 |
| 2025-10 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2025-11 | 30 | 26 | 4 | 0 | 0 | 0 |
| 2025-12 | 31 | 31 | 0 | 0 | 0 | 0 |
| 2026-01 | 31 | 29 | 2 | 0 | 0 | 0 |
| 2026-02 | 28 | 22 | 3 | 2 | 1 | 0 |
| 2026-03 | 31 | 25 | 2 | 4 | 0 | 0 |
| 2026-04 | 30 | 25 | 3 | 1 | 1 | 0 |
| 2026-05 | 31 | 26 | 1 | 3 | 1 | 0 |
| 2026-06 | 30 | 26 | 1 | 3 | 0 | 0 |
| 2026-07 | 31 | 24 | 4 | 3 | 0 | 0 |
| 2026-08 | 31 | 20 | 6 | 4 | 0 | 1 |
| 2026-09 | 30 | 19 | 9 | 1 | 1 | 0 |
| 2026-10 | 31 | 15 | 11 | 4 | 0 | 1 |
| 2026-11 | 5 | 3 | 2 | 0 | 0 | 0 |
Unused words in both the original and July 2023 lists are U_orig; July additions are U_2023. Previous answers are grouped by whether their first appearance was in 2021 or later. Other contains the remaining accepted guesses. Each puzzle is classified before its answer is observed.
What remains, and what repeats
U_2023: 785 unused words. At the last 90 puzzles' rate of 28 first uses, that would last about 2523 days. This assumes the rate stays constant; it is not a forecast.
U_orig: 440 unused words. At the last 90 puzzles' rate of 52 first uses, that would last about 762 days. This assumes the rate stays constant; it is not a forecast.
The first recorded repeat was puzzle #1689 on 2026-02-02.
- 25 repeated answers first appeared in 2021.
- 2 repeated answers first appeared in 2022.
- 1 repeated answers first appeared in 2023.
- 1 repeated answers first appeared in 2024.
What this means for solving
A captured answer list can exclude a valid answer. The answer-likelihood model assigns every captured accepted guess a positive weight and filters by the complete feedback history. Its estimates describe recorded category rates; we do not know how the NYT chooses answers. Published historical solver-pressure figures keep their original pool and are not rewritten to look like new-model results.
Method and limitations
The archive contains only puzzles completed worldwide. Source lists and both TARSE trees come from Alex Selby's pinned MIT-licensed repository. They are historical snapshots, not an official current exhaustive answer list. Counts use one row per puzzle; a repeat remains a separate row.
The model selects a 60-day half-life from six candidates using rolling-origin evaluation over the last 365 puzzles. Each fit uses only earlier answers, with a small positive prior. Model selection and the displayed scores use the same evaluation window, so these are not an independent estimate of future performance.
| Model | Mean log probability | Zero days | Median rank |
|---|---|---|---|
| v1-H30 | -7.1226 | 0 | 312.0 |
| v1-H45 | -7.1144 | 0 | 312.0 |
| v1-H60 | -7.1139 | 0 | 312.0 |
| v1-H90 | -7.1186 | 0 | 312.0 |
| v1-H180 | -7.1347 | 0 | 312.0 |
| v1-Hnone | -7.2263 | 0 | 312.0 |
| frozen-uniform-pool-retrospective | Negative infinity | 36 | 1177.5 |
Derived categories describe recorded sources, not NYT selection rules. A fixed-pool comparison is retrospective. Depletion assumes the recent first-use rate remains constant; repeat choices and editorial policy can change.
Reproduce this release
Run python scripts/refit_answer_model.py --as-of 2026-11-06. The cutoff is exclusive. The manifest records source checksums; the summary records the code release and implementation hash. The current model is separate from frozen published benchmarks.
Code release: fe3218e1b45fbba0c2c7abaffce750f7ccd37406.
Download and cite
These derived files omit answer words. Reuse them under CC BY 4.0; source word lists retain their own terms.
answer-sources.csv · answer-sources.json · schema.json · data-dictionary.json · summary.json · manifest.json · chart.svg · chart.png
Cite: FiveLetterWords. Where Wordle answers come from. Data through 2026-11-05. Release answer-sources-2026-11-05-07d29e34a580.