$ ~/aibattle/connect4
🔴 Connect Four · Perfect-information game · round-robin · 1974 games · board 6×7
🎬 Watch featured replays →
How Connect Four works
A two-player game on a vertical 7-column × 6-row grid; the two models alternate dropping discs of their own color. One game plays out like this:- On your turn you pick a column; the disc falls to the lowest empty cell in it. A full column can't be chosen.
- Player 0 moves first — a real edge in a solved game.
- The first to line up four of their discs in a row — horizontally, vertically, or diagonally — wins immediately.
- If the board fills with no four-in-a-row, the game is a draw.
Connect Four is solved: with perfect play the first player always wins. There's no hidden information or chance, so every position has a known best move — which lets us score tactical accuracy directly: did the model take an available immediate win, and did it block the opponent's immediate winning threat?
What the model sees each turn: the full current board, whose turn it is, and the legal (non-full) columns — perfect information.
1 · Results — who won
| # | Model | Elo | Net/game | Win% | 1st-move win% | 2nd-move win% | Games | tokens/dec | $/1K dec |
|---|---|---|---|---|---|---|---|---|---|
| 1 |  Claude Opus 4.8 |
1692 ±17 |
+0.49 | 73% | 75% | 72% | 493 | — | — |
| 2 |  GPT 5.5 |
1645 ±21 |
+0.30 | 63% | 61% | 66% | 269 | — | — |
| 3 | GPT 5.4 |
1630 ±23 |
+0.26 | 61% | 61% | 62% | 269 | — | — |
| 4 |  Kimi K2.6 |
1608 ±19 |
+0.22 | 59% | 63% | 55% | 346 | 18,188 | $72.75 |
| 5 |  DeepSeek V4 Pro |
1598 ±17 |
+0.21 | 59% | 62% | 57% | 378 | 11,686 | $40.67 |
| 6 |  GLM-5.2 |
1517 ±20 |
-0.01 | 48% | 53% | 43% | 300 | 16,151 | $71.06 |
| 7 | GLM-5.1 |
1503 ±18 |
-0.05 | 47% | 44% | 50% | 380 | 10,704 | $47.10 |
| 8 | Claude Sonnet 4.6 |
1434 ±17 |
-0.18 | 40% | 47% | 34% | 463 | — | — |
| 9 |  Qwen3.7 Plus |
1415 ±19 |
-0.29 | 35% | 39% | 30% | 300 | 10,572 | $16.92 |
| 10 |  MiniMax-M3 |
1350 ±22 |
-0.43 | 27% | 32% | 23% | 310 | 16,305 | $19.57 |
| 11 | MiniMax-M2.7 |
1306 ±34 |
-0.52 | 23% | 33% | 13% | 150 | 14,874 | $17.85 |
| 12 | GPT-OSS 120B |
1303 ±22 |
-0.50 | 24% | 32% | 17% | 290 | 3,157 | $1.89 |
columns: Elo opponent-adjusted rating · net/game mean result/game (+1/−1/0) · win% games won · 1st-move win% / 2nd-move win% win rate moving first / second · games games played (varies by wave)
Elo rating
Whiskers show ±1 bootstrap SD (resampling games 300×) — wider bars mean
fewer/less-decisive games, so ratings within a whisker of each other are a near tie.
Win / draw / loss
⚔️ Head-to-head (row wins–losses vs column)
| Claude Opus 4.8 | Claude Sonnet 4.6 | DeepSeek V4 Pro | GLM-5.1 | GLM-5.2 | GPT 5.4 | GPT 5.5 | GPT-OSS 120B | Kimi K2.6 | MiniMax-M2.7 | MiniMax-M3 | Qwen3.7 Plus | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 4.8 | — | 40-8 2d | 43-22 0d | 52-13 0d | 22-7 1d | 14-15 1d | 17-10 3d | 44-5 1d | 33-30 0d | 41-7 2d | 28-1 1d | 28-2 0d |
| Claude Sonnet 4.6 | 8-40 2d | — | 18-47 0d | 28-36 1d | 13-17 0d | 9-20 1d | 11-16 3d | 33-17 0d | 5-28 0d | 33-17 0d | 13-17 0d | 16-14 0d |
| DeepSeek V4 Pro | 22-43 0d | 47-18 0d | — | 17-11 2d | 22-8 0d | 16-12 1d | 11-18 0d | 25-5 0d | 11-16 3d | 10-0 0d | 24-5 1d | 19-9 2d |
| GLM-5.1 | 13-52 0d | 36-28 1d | 11-17 2d | — | 15-15 0d | 9-21 0d | 7-23 0d | 24-4 2d | 14-15 1d | 8-2 0d | 23-6 1d | 17-13 0d |
| GLM-5.2 | 7-22 1d | 17-13 0d | 8-22 0d | 15-15 0d | — | 11-17 2d | 9-19 2d | 24-6 0d | 17-10 3d | 0-0 0d | 19-10 1d | 17-12 1d |
| GPT 5.4 | 15-14 1d | 20-9 1d | 12-16 1d | 21-9 0d | 17-11 2d | — | 15-14 1d | 0-0 0d | 15-14 1d | 0-0 0d | 24-6 0d | 26-2 2d |
| GPT 5.5 | 10-17 3d | 16-11 3d | 18-11 0d | 23-7 0d | 19-9 2d | 14-15 1d | — | 0-0 0d | 15-14 1d | 0-0 0d | 27-3 0d | 28-2 0d |
| GPT-OSS 120B | 5-44 1d | 17-33 0d | 5-25 0d | 4-24 2d | 6-24 0d | 0-0 0d | 0-0 0d | — | 5-25 0d | 5-5 0d | 13-16 1d | 10-20 0d |
| Kimi K2.6 | 30-33 0d | 28-5 0d | 16-11 3d | 15-14 1d | 10-17 3d | 14-15 1d | 14-15 1d | 25-5 0d | — | 9-1 0d | 23-6 1d | 21-8 1d |
| MiniMax-M2.7 | 7-41 2d | 17-33 0d | 0-10 0d | 2-8 0d | 0-0 0d | 0-0 0d | 0-0 0d | 5-5 0d | 1-9 0d | — | 3-7 0d | 0-0 0d |
| MiniMax-M3 | 1-28 1d | 17-13 0d | 5-24 1d | 6-23 1d | 10-19 1d | 6-24 0d | 3-27 0d | 16-13 1d | 6-23 1d | 7-3 0d | — | 8-22 0d |
| Qwen3.7 Plus | 2-28 0d | 14-16 0d | 9-19 2d | 13-17 0d | 12-17 1d | 2-26 2d | 2-28 0d | 20-10 0d | 8-21 1d | 0-0 0d | 22-8 0d | — |
2 · Why — what decides win & loss
Tactical accuracy (win-take / block %)
Error rate by game phase
🛡️ Defense, not offense, decides games
Converting an available immediate win (win-take) is near-saturated for everyone —
it doesn't separate the field. The spread is all in defense (blocking the opponent's immediate
threat) and in late-game errors: every model plays the opening cleanly, so skill is just not
collapsing once threats pile up.
| model | win-take % | block % | late-game error % |
|---|---|---|---|
| Claude Opus 4.8 | 96 | 84 | 17 |
| Kimi K2.6 | 99 | 83 | 20 |
| DeepSeek V4 Pro | 99 | 79 | 21 |
| GLM-5.1 | 98 | 84 | 21 |
| GLM-5.2 | 94 | 83 | 25 |
| Claude Sonnet 4.6 | 100 | 76 | 27 |
| Qwen3.7 Plus | 96 | 74 | 33 |
| MiniMax-M3 | 96 | 75 | 34 |
| GPT-OSS 120B | 91 | 68 | 43 |
| MiniMax-M2.7 | 95 | 64 | 50 |
win-take clusters at 91–100% (offense solved); block% and late-game error % fan out across
the whole field and track the Elo order almost monotonically.
🎯 The decisive test: miss-rate by threat axis
Boards are sent as space-separated text, read row by row (
X . O . .). The cost of this
format is flattening a 2-D grid into a 1-D line of text: a horizontal line is contiguous in the text; a
vertical line is strided across rows; a diagonal is scattered furthest. Miss-rate climbs in exactly that
order. (Single, blockable immediate-loss threats only; unblockable double threats excluded.)| threat axis | faced | missed | miss-rate |
|---|---|---|---|
| horizontal (contiguous in text) | 825 | 78 | 9.5% |
| vertical (strided across rows) | 800 | 94 | 11.7% |
| diagonal ↘ (down-right) | 308 | 36 | 11.7% |
| diagonal ↙ (down-left) | 369 | 58 | 15.7% |
| diagonal (both) | 677 | 94 | 13.9% |
| all axes | 2302 | 266 | 11.6% |
Horizontal threats — laid out contiguously in the text — are the easiest to
defend; the further a line strays from the reading order, the more it is missed. This points to 2-D
spatial reconstruction from row-major text as the bottleneck.
↘ vs ↙ : why the two diagonals differ
The game is left-right symmetric, so an asymmetry between the two diagonals can't come
from the rules — it must come from the text representation. Following a diagonal means tracking a
row and a column index together: the ↘ main diagonal increments both indices in lockstep and flows
with the top-to-bottom, left-to-right reading order; the ↙ anti-diagonal increments the row
but decrements the column — two counters moving in opposite directions, against the reading
gradient — so it's missed more.
| model | ↘ down-right | ↙ down-left |
|---|---|---|
| Claude Opus 4.8 | 5% n=43 | 15% n=74 |
| GLM-5.1 | 9% n=46 | 9% n=57 |
| Claude Sonnet 4.6 | 7% n=46 | 11% n=45 |
| GPT-OSS 120B | 19% n=42 | 30% n=33 |
| Kimi K2.6 | 3% n=30 | 10% n=40 |
| DeepSeek V4 Pro | 18% n=33 | 15% n=27 |
| MiniMax-M2.7 | 19% n=26 | 34% n=29 |
| GLM-5.2 | 22% n=18 | 12% n=24 |
| Qwen3.7 Plus | 7% n=14 | 19% n=21 |
| MiniMax-M3 | 20% n=10 | 11% n=19 |
Pooled over both games: ↘ 13.3% vs ↙ 19.0% (z=2.65, p=0.004). The effect is weaker
and noisier here in Connect Four — gravity and short 4-lines make verticals/diagonals more salient — but
it is unanimous across all models in Gomoku, the larger gravity-free board.
3 · Analysis
Game-length distribution (plies)
🗺️ Move-location heatmap
Where each model places pieces (brighter = more frequent). Reveals
center-control bias and opening preferences.
Claude Opus 4.8
Claude Sonnet 4.6
DeepSeek V4 Pro
GLM-5.1
GLM-5.2
GPT 5.4
GPT 5.5
GPT-OSS 120B
Kimi K2.6
MiniMax-M2.7
MiniMax-M3
Qwen3.7 Plus