Hard mode on TicToe.org plays a mathematically perfect game. It uses the minimax algorithm to look at every possible move and pick the best one. In theory that means two things: you can never beat it, and you never have to lose to it either. Perfect play from both sides always ends in a draw.

The data from 727 real games against Hard mode says the first half of that theory is true. The second half is where things get interesting.

Not a single win

Across 727 completed games played by 86 different browsers, the Hard mode AI won 306 and drew 421. Players won zero.

Results against Hard mode (727 games):

The zero is expected. The 42% is not. Tic tac toe is a solved game, and the rules for never losing fit on a single page. Our own guide covers them. Yet more than two in five games against the AI ended in a loss.

The AI always opened in the same corner

When these games were played, the AI took the top-left corner every time it went first. In 334 games where the AI moved first, it never opened anywhere else. (It does now, more on that below.)

What happened next depended almost entirely on one square: where the player answered.

Player's reply to the AI's corner opening · 334 Hard mode games
Draw and loss rate for each reply to the AI's top-left opening, from 334 Hard mode games on TicToe.org. Only the centre ever drew.

Answering in the centre drew 118 of 139 games. Answering anywhere else lost 195 out of 195.

This is the one rule in tic tac toe that has no exceptions. If your opponent opens in a corner, you must take the centre. Any other reply hands a perfect opponent a forced win, and Hard mode never misses one.

The trap behind most losses

The single most common game in the entire dataset is a loss, and it was played 37 times move for move.

It goes like this. The AI takes the top-left corner. The player answers in the top-right corner, which feels natural because corners are strong. The AI then takes the middle-left square, threatening the left column. The player has to block in the bottom-left corner. Now the AI takes the centre, which creates two threats at once: the diagonal and the middle row. The player can only block one. The AI wins on its fourth move.

That double threat is called a fork. Against a perfect player you only need to slip once, by missing a block or by giving it the square it needs to set up a fork, and the game is over. In this game the mistake happened on the player's very first move.

Most losses were reruns

Here is the part that surprised us most.

Those 306 losses came from just 110 different games. The other 196 losses, 64% of the total, were exact repeats of a losing game someone had already played. The five most common losing games alone account for 112 losses.

The five most repeated losses:

At the time, Hard mode was deterministic: the same position always got the same answer. Lose once, play the same moves again, and you lost in exactly the same way. One player lost the identical game five times.

That was the flip side of a perfect opponent that never varied. It could not be beaten, but it could be learned completely. Every loss was a recipe you could read and avoid.

Players got worse, not better

You might expect players to learn as they go. The data says the opposite.

We looked at the 22 browsers that played Hard mode at least ten times and split each one's games in half. In the first half of their games they lost 41% of the time. In the second half they lost 45%.

The most likely explanation is tilt. Players who keep losing keep playing, often faster, and fall back on the same opening that lost them the previous game. Against a human that might occasionally work. Against Hard mode it never does.

When you go first, the centre is safest

When the player moves first, the AI cannot win unless the player makes a mistake. Players drew 303 and lost 90 of their 393 games.

Loss rate by the player's opening square (393 games):

Middle-left and middle-right were opened too rarely (6 and 8 games) to say anything useful about.

This is a neat contrast with our analysis of 50,008 online games, where the centre had the lowest win rate. Against a human, the centre draws a lot because it attracts careful opponents. Against a perfect opponent, the centre is simply the opening that leaves you the fewest ways to go wrong.

What this means if you are playing Hard mode

Most important: If the AI opens in a corner, take the centre. Every other reply in our data lost, all 195 times. Any corner works the same way, since the board is symmetrical.

If the AI opens in the centre: take a corner. An edge reply gives a perfect player room to set up a fork.

If the AI opens on an edge: the centre is always a safe reply.

If you go first: open in the centre. It lost only 11% of the time, the lowest of any opening with a meaningful sample.

After every move: check whether the AI has two in a row before you do anything else. Then check whether your move leaves it a square that creates two threats at once. Those two checks are the whole game.

A draw is the best possible result against Hard mode, and it is always available. Try Hard mode and see whether you can go a full session without a loss. If you want to understand how the AI decides, our explanation of the minimax algorithm walks through it step by step.

Why Hard mode now varies its moves

After seeing this data, we changed how Hard mode chooses. It still plays perfectly and still cannot be beaten. But when several moves are equally good, it now picks one of them at random instead of always taking the first.

In tic tac toe that happens a lot. Every opening square leads to a draw with perfect play, so the AI can now open anywhere on the board. Many later positions also have more than one safe reply.

The practical effect is that the reruns are over. You can no longer memorise one losing game and replay it, and you cannot learn a single winning path for the AI either, because there is no longer only one. What stays exactly the same is how you lose: by walking into a fork. The rules for avoiding one have not changed.

We will compare the loss rate before and after the change once enough new games have come in.

How this compares to Easy mode and online play

Easy mode is a different story entirely. Players won 172 of 238 games against it, 72%. In online games between two people, 58% of games ended in a draw, the same share as against Hard mode. The difference is that against each other, the remaining games split roughly evenly between the two players. Against Hard mode, all of them went one way.

About this data

All figures are drawn from completed games against the Hard mode AI on TicToe.org between 9 and 11 October 2026. The dataset covers 727 games from 86 different browsers. Abandoned games are not included. All of these games were played before Hard mode started choosing at random between equally good moves, so the AI's opening and replies were fixed throughout.

The same limitation applies as in our other analyses: we cannot guarantee that every game comes from a unique player. A few enthusiastic players contribute many games each, and a browser is not the same as a person. The patterns should be treated as observational rather than experimental. The sample is also small compared to our online multiplayer data, so the opening loss rates for rarely chosen squares should be read with caution.

Data collection is ongoing. This post will be updated as more games accumulate.

Sources