Connect Four table

Shelved, unfinished. 37 of 100 games were played on 9 October 2026, then the table was stopped: the full set would have cost about $25. Every square here is a real game, the empty squares were never played, and 3 of the 8 models never got a game. The one thing to take from it: the random player won 2 of its 10 games and the perfect player won 10 of its 10 games, and the models sit in between.

The table

The model on the left moved first; the model along the top moved second. W means the model on the left won, L means it lost, D is a draw. The small number is how many moves the game took.

• First player won
• Second player won
• Draw

Standings

In no particular order, because the schedule was never finished: a player with three games cannot be ranked against one with eight. Games a model played against itself are on the grid but left out of these counts. Two yardsticks play in the same table: a player that picks columns at random, and a solver that never makes a mistake.

• Best move is how often the model played a move the solver rates as good as any other.
• A blunder is a move that threw away a won game or turned a drawn one into a loss.
• A missed win is a turn where four in a row was there to take and the model played somewhere else.
• Thinking a move is the average number of tokens the model spent reasoning before it answered.

How it works

• Every model plays every other model twice, once moving first and once second, and plays itself once. One game per square.
• For each move the model gets one message: the rules, the board, the columns played so far and which piece it is. It has no memory of earlier moves, no other instructions and no tools.
• It must answer with a column number. An answer that cannot be played gets one more try, then the model forfeits the game.
• Every model was asked for as little thinking as it allows. Some think anyway; the table shows how much.
• DeepSeek ignores minimal and thinks until it runs out of room (it forfeited 9 of 12 sample games that way), so its thinking is switched off, which it allows.
• Models are reached through OpenRouter, each on its maker's own servers. This run cost $2.53 in all, counting retries and answers that could not be played.

The message a model gets

You are playing Connect Four.

The board has 7 columns and 6 rows. Two players take turns. On your turn you choose a column and your piece drops to the lowest empty space in it. The first player to get four of their own pieces in a line, across, up and down or diagonally, wins. If the board fills up first, the game is a draw.

You are {you}. Your opponent is {them}. {first} moved first. It is your turn.

The board, top row first. A dot is an empty space. The numbers are the columns.

{board}

Columns played so far, in order: {moves}

Reply with the number of the column you choose, 1 to 7, and nothing else.

What this does not show

• One game per square is a small sample. A single result can be luck; the totals mean more than any one square.
• With thinking turned down, this measures quick play, not a model's best effort.
• Connect Four is a solved game. Moving first wins with perfect play, so the first player has an edge.
• What each model said was stored cut to 200 characters, so a replay shows the start of its answer, not all of it. The reading of answers was not identical for every model. Both would be fixed before any further run.

All the tests