$ ~/aibattle/gomoku
⚫ Gomoku-Lite · Perfect-information game · round-robin · 1957 games · board 9×9
🎬 Watch featured replays →
How Gomoku-Lite works
A two-player game on a 9×9 board (columns A–I, rows 1–9). The two models alternate placing one stone of their own color on any empty cell:- On your turn, name an empty cell — e.g.
E5— and a stone is placed there. - Player 0 moves first.
- The first to make five of their stones in a row — horizontally, vertically, or diagonally — wins immediately.
- If the board fills with no five-in-a-row, the game is a draw (rare on 9×9).
Free-style Gomoku — no forbidden-move restrictions — and like Connect Four there's no hidden information or chance, so we score tactical accuracy: did the model complete an available five (win-take) and block the opponent's immediate five-threat (block rate)?
What the model sees each turn: the full current board, whose turn it is, and that any empty cell is a legal move — perfect information.
1 · Results — who won
| # | Model | Elo | Net/game | Win% | 1st-move win% | 2nd-move win% | Games | tokens/dec | $/1K dec |
|---|---|---|---|---|---|---|---|---|---|
| 1 |  GPT 5.5 |
1686 ±22 |
+0.38 | 68% | 83% | 53% | 262 | — | — |
| 2 |  Kimi K2.6 |
1657 ±21 |
+0.37 | 69% | 78% | 59% | 345 | 16,740 | $66.96 |
| 3 |  GLM-5.1 |
1603 ±18 |
+0.23 | 61% | 64% | 58% | 379 | 8,268 | $36.38 |
| 4 | GLM-5.2 |
1596 ±19 |
+0.19 | 59% | 65% | 53% | 299 | 16,574 | $72.93 |
| 5 | GPT 5.4 |
1593 ±22 |
+0.12 | 55% | 65% | 45% | 257 | — | — |
| 6 |  DeepSeek V4 Pro |
1573 ±19 |
+0.16 | 58% | 63% | 52% | 363 | 9,888 | $34.41 |
| 7 |  Claude Opus 4.8 |
1529 ±16 |
+0.08 | 53% | 67% | 42% | 464 | — | — |
| 8 | Claude Sonnet 4.6 |
1510 ±17 |
+0.00 | 50% | 58% | 43% | 495 | — | — |
| 9 |  Qwen3.7 Plus |
1444 ±20 |
-0.24 | 38% | 43% | 32% | 300 | 11,609 | $18.57 |
| 10 |  MiniMax-M3 |
1344 ±22 |
-0.46 | 27% | 35% | 19% | 310 | 13,510 | $16.21 |
| 11 | GPT-OSS 120B |
1243 ±28 |
-0.63 | 18% | 24% | 12% | 290 | 2,420 | $1.45 |
| 12 | MiniMax-M2.7 |
1220 ±37 |
-0.65 | 17% | 25% | 9% | 150 | 13,232 | $15.88 |
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 | — | 24-26 0d | 27-38 0d | 26-39 0d | 22-8 0d | 16-14 0d | 8-20 2d | 39-11 0d | 6-28 0d | 43-7 0d | 23-7 0d | 14-14 2d |
| Claude Sonnet 4.6 | 26-24 0d | — | 25-40 0d | 31-34 0d | 12-17 1d | 14-15 1d | 4-26 0d | 41-8 1d | 15-50 0d | 38-12 0d | 20-10 0d | 20-10 0d |
| DeepSeek V4 Pro | 38-27 0d | 40-25 0d | — | 14-16 0d | 10-20 0d | 9-10 0d | 8-16 0d | 24-6 0d | 15-15 0d | 9-1 0d | 21-9 0d | 22-8 0d |
| GLM-5.1 | 39-26 0d | 34-31 0d | 16-14 0d | — | 12-17 0d | 15-15 0d | 12-17 1d | 29-1 0d | 14-16 0d | 9-1 0d | 27-3 0d | 25-5 0d |
| GLM-5.2 | 8-22 0d | 17-12 1d | 20-10 0d | 17-12 0d | — | 13-16 1d | 8-21 1d | 28-2 0d | 14-16 0d | 0-0 0d | 27-3 0d | 24-6 0d |
| GPT 5.4 | 14-16 0d | 15-14 1d | 10-9 0d | 15-15 0d | 16-13 1d | — | 14-16 0d | 0-0 0d | 10-18 0d | 0-0 0d | 24-6 0d | 24-5 1d |
| GPT 5.5 | 20-8 2d | 26-4 0d | 16-8 0d | 17-12 1d | 21-8 1d | 16-14 0d | — | 0-0 0d | 18-10 0d | 0-0 0d | 23-7 0d | 22-8 0d |
| GPT-OSS 120B | 11-39 0d | 8-41 1d | 6-24 0d | 1-29 0d | 2-28 0d | 0-0 0d | 0-0 0d | — | 1-29 0d | 6-4 0d | 12-18 0d | 6-24 0d |
| Kimi K2.6 | 28-6 0d | 50-15 0d | 15-15 0d | 16-14 0d | 16-14 0d | 18-10 0d | 10-18 0d | 29-1 0d | — | 10-0 0d | 26-4 0d | 19-11 0d |
| MiniMax-M2.7 | 7-43 0d | 12-38 0d | 1-9 0d | 1-9 0d | 0-0 0d | 0-0 0d | 0-0 0d | 4-6 0d | 0-10 0d | — | 1-9 0d | 0-0 0d |
| MiniMax-M3 | 7-23 0d | 10-20 0d | 9-21 0d | 3-27 0d | 3-27 0d | 6-24 0d | 7-23 0d | 18-12 0d | 4-26 0d | 9-1 0d | — | 8-22 0d |
| Qwen3.7 Plus | 14-14 2d | 10-20 0d | 8-22 0d | 5-25 0d | 6-24 0d | 5-24 1d | 8-22 0d | 24-6 0d | 11-19 0d | 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
Completing an available five (win-take) is near-saturated for everyone — it doesn't
separate the field. The spread is all in defense (blocking the opponent's immediate five-threat)
and in late-game errors: every model plays the opening cleanly, so skill is just not collapsing
once threats pile up. Gomoku's bigger 9×9 board makes defense harder, so block% sits lower than in
Connect Four across the board.
| model | win-take % | block % | late-game error % |
|---|---|---|---|
| GLM-5.2 | 92 | 70 | 24 |
| Kimi K2.6 | 100 | 74 | 19 |
| GLM-5.1 | 94 | 67 | 31 |
| DeepSeek V4 Pro | 100 | 63 | 32 |
| Claude Sonnet 4.6 | 97 | 61 | 41 |
| Claude Opus 4.8 | 90 | 58 | 43 |
| Qwen3.7 Plus | 93 | 53 | 51 |
| MiniMax-M3 | 92 | 49 | 60 |
| GPT-OSS 120B | 86 | 47 | 65 |
| MiniMax-M2.7 | 90 | 47 | 70 |
win-take clusters at 86–100% (offense solved); block% and late-game error % fan out across
the whole field and track the Elo order almost monotonically. Claude is the standout anomaly — strong in
Connect Four but its late-game error rate balloons to ~43% here on the sparse 9×9 board.
🎯 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) | 428 | 24 | 5.6% |
| vertical (strided across rows) | 401 | 49 | 12.2% |
| diagonal ↘ (down-right) | 284 | 43 | 15.1% |
| diagonal ↙ (down-left) | 231 | 56 | 24.2% |
| diagonal (both) | 515 | 99 | 19.2% |
| all axes | 1344 | 172 | 12.8% |
Horizontal threats — laid out contiguously in the text — are the easiest to
defend (5.6%), while ↙ diagonals are missed 4× as often (24.2%). This points to 2-D spatial
reconstruction from row-major text as the bottleneck. The effect is sharper here than in Connect Four:
the 9×9 gravity-free board leans entirely on reading lines out of the text.
↘ 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 | 21% n=63 | 34% n=47 |
| Claude Sonnet 4.6 | 10% n=52 | 12% n=43 |
| DeepSeek V4 Pro | 6% n=32 | 25% n=28 |
| GPT-OSS 120B | 23% n=30 | 30% n=23 |
| MiniMax-M2.7 | 15% n=27 | 29% n=24 |
| Kimi K2.6 | 16% n=32 | 26% n=19 |
| GLM-5.1 | 12% n=25 | 14% n=21 |
| MiniMax-M3 | 20% n=10 | 42% n=12 |
Pooled over both games: ↘ 13.3% vs ↙ 19.0% (z=2.65, p=0.004). The direction is
unanimous in Gomoku — every one of the eight models above misses ↙ more than ↘ — which is strong
evidence the asymmetry is systematic, not noise.
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