How accurate is it, really?
The strongest published crossword AIs are Berkeley's solver (~99.9% letters, ~82% full grids) and Dr.Fill, which beat humans at the 2021 American Crossword Puzzle Tournament. Here is where we stand — measured the honest way: on held-out public-domain puzzles, with memorization disabled, so the number reflects reasoning, not recall.
How we compare
| System | Letters | Full grids |
|---|---|---|
| Across & Down (this engine) | 99.9% | 86.7% |
| Berkeley Crossword Solver (2022) | ~99.9% | ~82% |
| Dr.Fill (ACPT 2021) | ~99%+ | ~superhuman |
| Typical "crossword solver" sites (pattern match) | n/a | ~0% |
Methodology (so you can trust the number)
- · Held-out & honest: every clue's exact-answer memory is switched off (
proofMode), so the engine must reason, not recall. - · Public-domain only: pre-1965 NYT puzzles (xd.saul.pw archive). No copyrighted material.
- · Letter accuracy = % of white squares filled correctly. Full grids = % of puzzles solved 100% correct.
- · Reproducible: the engine and corpus are the same ones you can use right now under Solve.
- · Sample size: 30 puzzles, run 6/20/2026. We grow this over time and show the latest.
Per-puzzle results
green = perfect grid · amber = ≥95% letters · red = below. Hover for detail.
Honest caveat — read this: our held-out set is pre-1965 public-domain puzzles, which are gentler than the modern, theme-heavy grids Berkeley and Dr.Fill were measured on. So treat this as "same league on a fair, honest test," not a head-to-head win — we won't claim a crown we haven't earned on identical puzzles. Closing that gap (modern puzzles, harder themes) is exactly what we build in the open. Watch a fresh solve every day on the Daily Proof.