How we tested Next Tile

We check Next Tile's picks against a slower, deeper search, and its engine against a second one. Here's what we found, as of 2026-10-29.

Results

It picks what a much deeper search picks. On 12 positions from real Crossplay games, we compared its move with a search that tries 20 moves and plays each one out 600 times. They agreed 10 times. The other two were too close to call, and on average Next Tile gave up 0.2 points.
It solves endgames exactly. With one tile left in Crossplay, it checks every play, every tile you might draw and every reply. Once the bag is empty in Scrabble GO, it plays every line out to the end of the game, and uses a shorter search only if that would take more than 10 seconds. Tests check both against a plain search of every move.
It's fast enough to think ahead. The engine is written in Rust. Playing a move out 240 times takes 0.6 seconds instead of 19, so a turn takes 3 to 7 seconds and close calls get played out up to 480 times.
Reading their last move sharpens its guesses. Working out which tiles your opponent kept from their last move makes its guesses about their rack 6 to 10% more accurate, measured over thousands of computer-vs-computer games.
The screenshot reader checks itself. Every word it reads is checked against the word list. When a letter is unclear, it only swaps in its second guess if that makes every word on the board valid, and it marks the square for you.
Speed never changes the answer. The fast engine is checked against a slower reference version on every move, score and tile value.