Strategy guides › 2048
How to reach 2048: why the corner rule alone fails
▶ Play 2048🎯 Play the seed our bot took to 4096
Your 512 sits in the bottom-left corner, a 256 beside it, a 128 next to that. One more 128 and the whole row collapses into a 1024. Then a 2 lands in the one gap you needed, you press left out of habit, and the 512 slides out of its corner for good.
Everyone has heard the tip: keep your biggest tile in a corner. It is true, but on its own it is nowhere near enough. We measured it.
What we tested
We played thousands of games with the game’s own code (the same slide-and-merge function and the same tile spawner the page runs) using four kinds of player. A random player that presses any direction that moves something. A corner player that always presses down if it can, otherwise left, then right, and up only when nothing else moves. And two look-ahead players that, before each swipe, imagine where the next tile could land and score the result, preferring empty squares and tiles that grow towards the bottom-left corner. One looks a single move ahead; the other looks two moves ahead (your move, the new tile, your next move, the next tile). We ran 1,000 games each for the random and corner players, 500 for one move ahead and 300 for two.
| Player (4×4) | Median score | Made 1024 | Made 2048 |
|---|---|---|---|
| Random | 1,028 | 0% | 0% |
| Corner rule | 2,420 | 0% | 0% |
| One move ahead | 3,856 | 0% | 0% |
| Two moves ahead | 22,364 | 93% | 50% |
The corner rule more than doubles a random player’s score, and then hits a wall: in a thousand games it made a 512 only 3.7% of the time and never a 1024. Looking one move ahead lifts the 512 rate to 18%, still with no 1024. Looking two moves ahead, the same corner idea makes 2048 in half of all games, and 4096 in 3%. Searching one move deeper still is slow in a browser-speed script, but in the 12 games we let it play it made 2048 eight times and 4096 six times.
Order, not just a corner
Why does a little foresight help so much? Because the corner is only the anchor. What actually wins is the shape of the whole board: big tiles in the corner, every row getting smaller away from it, every column getting smaller as it goes up. Then each tile sits next to the one it will eventually merge with, and a merge in one place sets off the next.
We checked how much each ingredient is worth by taking pieces out of the two-move player and replaying 150 games each. Counting only empty squares, it never made 2048 (1024 in 45% of games, median score 7,384). Adding a preference for neighbouring tiles of similar size lifted that to 2048 in 8%. Adding the corner-and-order preference took it to 50%. The corner matters, but only as part of an ordered board, and only if you check where the next tile can land before you swipe.
The question to ask before every swipe
You cannot run a search in your head every move, but you can ask the question the search asks: after this swipe, where can the new tile appear, and will I still have a good move then?
The new tile appears on a random empty square: a 2 nine times in ten and a 4 one time in ten, the same odds as the original game (across more than 300,000 spawns in our two-move games, 10.0% were 4s). Any empty square is equally likely, so a swipe that leaves only one or two gaps tells you almost exactly where the next tile will land, and a swipe that leaves a gap inside your ordered rows invites a small tile into the middle of your chain. That is what the look-ahead player is weighing, and it is why counting empty squares alone was worth so much in the test above.
The look-ahead player is not religious about directions, either. Across 20 of its games it pressed left 42% of the time, down 42%, and up and right only about 8% each. It avoids the two “away” directions, but it uses them when the board needs them, rather than as a last resort.
Board sizes change the game
Our 3×3 and 5×5 boards are not just harder and easier versions of 4×4.
On 3×3, 2048 is impossible. To build a tile you must at some point hold its two halves at once, and to build the second half you must hold its two halves, and so on down to a 4, which can arrive straight from the spawner. So a 1024 needs 512, 256, 128, 64, 32, 16, 8 and two 4s on the board together: nine tiles, nine squares, and a 2048 would need a tenth. (The same counting gives 131,072 as the largest tile possible on 4×4, which matches the known limit.) In practice even 1024 is out of reach; our two-move player topped out at 256, which it reached in 36% of 300 games. On 3×3, play for score, not for the tile.
On 5×5, the extra nine squares forgive almost anything. Even the random player reached 1024 in 23% of 1,000 games and 2048 in 0.6%; the plain corner rule made 2048 in 6.5%. Our two-move player made 2048 in all 60 games we gave it, and 16,384 in 29 of them. Here the corner habit pays off on its own, because there is room to recover from the moves it gets wrong.
Practice
The practice button opens seed 1003 on 4×4: the game our two-move player took all the way to 4096, so a 2048 is definitely in it. The spawns follow the seed, so with the same moves you would get the same tiles. Build the shape from the diagram early — big tiles bottom-left, everything shrinking away from them — and before each swipe, find the square where the next tile could land.
The corner is where you keep your big tile; the order is what keeps it company.
