We Tested 2048 Strategies by Computer: What the Corner Strategy Is Really Worth
Short answer: We let six computer players play 1,000 games of 2048 each. Pressing the keys in a circle scored no better than random moves. A player that blindly follows the corner rule (down, then left, then right, up only when forced) doubled the random score but never made a 1024 tile. A player that keeps the corner and also thinks about where the next tile can land reached 1024 in 87% of games and 2048 in 36%. The same player looking just as far ahead but ignoring the new tile was no better than one that looks only one move ahead. The corner is a good habit; what wins is thinking about the next tile.
- 2048: what the corner strategy is really worth (you are here)
- Pairs: how many tries does a perfect memory need?
- Code Breaker: how many guesses does it take?
"Keep your biggest tile in the corner" is the most repeated piece of 2048 advice, ours included: our guide to winning 2048 is built around it. Advice like this is easy to repeat and hard to check, because nobody plays a thousand games the same way. A computer can. So we wrote a small program that plays our 2048 with exactly the same rules, and gave it six different ways of choosing a move.
Six players, 1,000 games each

- Random: any move that changes the board.
- Circle: down, left, right, up, down, left… in turn. A well-known "trick" that is supposed to get you far without thinking.
- Blind corner: the corner rule as a reflex. Press down if it changes the board, otherwise left, otherwise right, and up only if nothing else works.
- One move ahead: tries each of the four moves and picks the one that leaves the best-looking board: many free cells, tiles getting smaller away from the bottom-left corner, and the biggest tile in that corner.
- Two moves, no new tile: the same judgement, but it plans two of its own moves ahead as if no new tile would appear in between.
- Two moves with the new tile: the same again, but for every move it also considers every empty cell where the new tile could appear, as a 2 or a 4, and its own best reply to each. It picks the move that leaves the best board on average.
The last three players share the same scoring rule, which we tuned by hand. Different weights would give somewhat different numbers, but comparing the three shows what looking ahead is worth. Game 1 of every player starts from the same board with the same stream of random numbers, game 2 from another, and so on; once the players move differently, their boards and new tiles go their own way.
The results
| Player | Median score | Reached 256 | Reached 512 | Reached 1024 | Reached 2048 |
|---|---|---|---|---|---|
| Random | 1,052 | 6% | 0% | 0% | 0% |
| Circle | 1,060 | 8% | 0.1% | 0% | 0% |
| Blind corner | 2,396 | 49% | 4.8% | 0% | 0% |
| One move ahead | 3,376 | 79% | 8.0% | 0% | 0% |
| Two moves, no new tile | 2,756 | 66% | 15% | 0.2% | 0% |
| Two moves with the new tile | 17,768 | 99.9% | 98.9% | 86.7% | 35.6% |
1,000 games per player. "Reached" means the tile appeared at least once during the game.
1. Pressing keys in a circle does nothing
The circle scored 1,060 against 1,052 for random moves. It feels as if it works, because the board keeps moving and merges happen, but it pays no attention to where the big tiles end up. Up is a quarter of its moves, and about half of those pull the bottom-left tile out of its place.
2. The corner rule helps, but on its own it stops at 512
Blindly following the corner rule more than doubles the score, and the biggest tile really does stay put: after 71% of moves it was in the bottom-left cell. Yet in 1,000 games it never made a 1024 tile. Its weakness is that it never compares moves. It presses down whenever down is possible, whatever that does to the rest of the board, and when it is finally forced to press up, every one of those up moves shifts its corner tile.
3. Looking further ahead only helps if you think about the new tile
This is the clearest result of the test. Planning two moves ahead while ignoring the new tile is no better than looking one move ahead: a lower median score, and it still almost never reached 1024. It builds plans that the next random tile ruins, and it kept its biggest tile in the corner after only 36% of moves. Add one question to the same plan, where can the next tile land, and what will I do then?, and the 2048 tile goes from never to more than one game in three.
The strong player does not simply avoid risky-looking boards. It actually left an empty cell in the bottom row slightly more often than the one-move player (after 28% of moves against 23%). The difference is that before choosing, it checks what a new tile in each empty cell would do and how well it could answer, and picks the move whose outcomes are best on average.
4. The strong player presses up, but almost never moves its corner tile
The strong player pressed up a median of 83 times per game, roughly one move in thirteen. Only 3% of those up moves shifted the tile in the bottom-left corner. They were not harmless, though: in 90% of the others, some other tile in the bottom row was lifted or merged upwards. The strong player can afford that because it has already checked what the new tile can do to the gaps it leaves. The blind corner player pressed up only when it had no other move, and by then the board was so packed that its corner tile always moved.
So "never press up" is a simplification of what good play looks like. The corner tile is the one thing the strong player almost never gives up; the rest of the bottom row it uses more freely, but never without looking.
How long a winning game takes
In the games where the strong player made a 2048 tile, it got there after a median of 1,050 moves, with a score of about 21,500 at that moment, a little above the minimum worked out in our 2048 guide, because a real board also holds a 1024, a 512 and smaller tiles waiting to merge. If you are at 20,000 points and still have no 2048, you are in the normal range.
How far can a computer go?
Our strong player is deliberately simple; it plays a whole game in about a tenth of a second. Serious 2048 programs search much deeper. A well-known open-source one by Robert Xiao, searching up to eight moves ahead, reached the 16,384 tile in 94 of 100 games and 32,768 in about a third of them. On a 4 × 4 board the largest tile that can exist at all is 131,072, and only if the game keeps handing you 4s at the right moments.
What this means when you play
The first two tips follow directly from the table. The others are our reading of how the strong player behaves, not separately measured rules.
- Keep the corner, but think before each move. A fixed key pattern, even a good one, tops out around 512.
- Before a move, look at the empty cells it will leave and ask what a new 2 would do in each. That one question is the difference between the two "two moves" players.
- If you must press up, make sure your corner column cannot move. A full corner column with no two equal tiles next to each other stays put. The other bottom-row tiles may still lift, so check what can land in the gaps (tip 2).
- A gap in the bottom row is acceptable if every tile that could land there can be merged away. If a 2 in that gap would be stuck between big tiles, choose another move.
- Slow down at the critical moments. Most moves are easy. The ones where your corner tile could move deserve a few seconds of thought.
Did you know?
- 2048 was written by the Italian developer Gabriele Cirulli over a weekend and released on 9 March 2014. He described it as a clone of the app 1024, which itself followed the iOS game Threes.
- Building this test uncovered a bug in our own game. In a row like 4-4-2-2, our 2048 merged only the first pair (giving 8-2-2) instead of both (8-4). The test program used the correct rule from the start, and the game has been fixed.
If you prefer a game with no luck at all, try Sudoku; for deduction under a guess limit, see Code Breaker.
Next: Part 2: Pairs, how many tries a perfect memory needs, and what forgetting costs.
