A Self-Play Study of 9x8 Connect Four: 1,395 Games, Every Opening Mapped
9x8 Connect Four is unsolved - nobody has even counted its positions. Until somebody does, the honest way to learn about the board is to measure. This page reports 1,395 games of seeded self-play between the Champion engine's strength levels - including, through a research-only port, the strongest one the Android app ships, now level 10. Everything here is Engine evidence - the measured behaviour of one learned evaluator, reproducible from the published seeds, and no substitute for perfect play.
Key findings
- 9 / 9 the second player has a winning reply to every one of the nine openings.
- 51 : 29 but careless replies flip it: the first player takes 51 of the 81 forced openings.
- 141 : 21 Max (the app's MCTS level) beats the Strong engine line from either seat.
- 54% / 56% parity echo: first-player wins land on odd rows, second-player wins on even rows.
- ~50:50 equal-strength mirrors show no significant first-move advantage.
- ply ~35 the evaluator's first-player optimism decays to dead even as the board fills.
- 1,395 seeded, reproducible games - full move lists downloadable below.
Method, in one paragraph
Engine All games use the shipped engine exactly as it plays: tactical rails (an immediate win is always taken, an immediate loss always blocked), the opening book, and the strength levels below. Randomness comes only from a seeded generator (base seed 20260803), so every game replays identically. Advanced - the in-game Strong - and Grandmaster are deterministic, so their variety comes from exhaustive openings instead of noise: we force each of the 81 possible two-ply starts and play them out. Grandmaster joins through a research-only port of the app engine, verified to reproduce the real engine's deterministic self-play line move for move (59 of 59 plies). Nothing about training is used or revealed - this is the engine as any player meets it. The study uses the engine's internal level names; this is how they map to the names in the game:
Note added 11 August 2026. This study was run on the four-level engine that shipped until August 2026. The game now offers ten levels grouped in the same four tiers, and the four engines measured here are unchanged - they are now levels 4 (Beginner), 6 (Intermediate), 8 (Advanced) and 10 (Grandmaster), with three genuinely beatable levels added below them and three interpolations in between. Every number and every downloadable file on this page is left exactly as it was measured; the Champion AI page carries the current ladder.
| In the game (until Aug 2026) | In this study | Now level | What it is | Included? |
|---|---|---|---|---|
| Easy | Beginner | 4 | Tactical rails plus a local heuristic with softmax sampling (temperature 0.28) - no neural network. | Yes |
| Medium | Intermediate | 6 | The network's move suggestions blended with the safe heuristic, softmax 0.15. | Yes |
| Strong | Advanced | 8 | The real model - the full engine line: opening book plus value-guided search over the neural network. Deterministic. | Yes - it plays the entire opening map |
| Max | Grandmaster | 10 | A 256-simulation Monte Carlo tree search over the same network; it topped the four-level ladder at 1798 against Beginner's 1000, and tops the current ten-level ladder too. | Yes - through a research-only port of the app engine, parity-checked move for move against the real thing. Playable only in the Android app. |
The opening map: all 81 starts, played out
A note before reading it: this 9-by-9 grid is not the board (the board is 9x8). It is the space of openings - the first player has nine possible first moves and the second player nine possible replies, so there are 9 x 9 = 81 two-ply starts, and each cell is one deterministic Strong-vs-Strong playout of that start. Green cells: the first player wins. Red: the second player. Gray: draw (there is exactly one - first move column 9, reply column 3). The number in each cell is the game length in plies - an empirical distance-to-win map:
Read column by column and the sharper fact appears. For every first move, at least one reply wins for the second player:
So at engine level, 9x8 behaves like a game where the reply matters more than the opening: against best cached play the second player always had a winning answer (51 : 29 for the first player only counts careless replies). Two honest footnotes. First, the engine's own book line ends in a 7-ply first-player win - the book's cached reply (answering straight on top) walks into a bottom-row double threat that two other replies avoid and refute, so the cache contains its own refutation. Second, the map is not mirror-symmetric (51 of 81 cells match their mirror, 30 do not): mirror positions are objectively identical, so this asymmetry measures the evaluator's own bias amplified by deterministic play, not a property of the board.
The strength ladder, round-robin
200 games per pairing (Grandmaster pairings 100 and 50), seats swapped halfway, Beginner and Intermediate sampling with their shipped temperatures:
| First player vs second | Games | P1 wins | Draws | P2 wins | P1 win rate | Mean plies |
|---|---|---|---|---|---|---|
| Beginner vs Beginner | 200 | 90 | 11 | 99 | 45.0% | 47.0 |
| Intermediate vs Intermediate | 200 | 101 | 10 | 89 | 50.5% | 50.2 |
| Beginner vs Intermediate | 100 | 9 | 1 | 90 | 9.0% | 39.8 |
| Intermediate vs Beginner | 100 | 88 | 1 | 11 | 88.0% | 41.5 |
| Beginner vs Advanced | 100 | 4 | 0 | 96 | 4.0% | 28.3 |
| Advanced vs Beginner | 100 | 100 | 0 | 0 | 100.0% | 25.9 |
| Intermediate vs Advanced | 100 | 28 | 0 | 72 | 28.0% | 28.3 |
| Advanced vs Intermediate | 100 | 97 | 0 | 3 | 97.0% | 34.4 |
| Grandmaster (Max) vs Intermediate | 50 | 48 | 0 | 2 | 96.0% | 35.9 |
| Intermediate vs Grandmaster (Max) | 50 | 0 | 0 | 50 | 0.0% | 34.8 |
| Grandmaster (Max) vs Beginner | 25 | 25 | 0 | 0 | 100.0% | 27.3 |
| Beginner vs Grandmaster (Max) | 25 | 0 | 0 | 25 | 0.0% | 25.7 |
Three readings. In equal-strength mirrors the first move is worth little: Beginner mirrors went 45.0% for the seat, Intermediate mirrors 50.5% - both within the roughly ±7 percentage points a 200-game sample allows, so no significant seat advantage. Skill dominates the seat one rung apart: Intermediate beats Beginner about 89% from either chair, Advanced concedes zero games to Beginner as first player, and Max drops just 2 of 150 games against the sampling levels - none as second player. But in the closest contest below the top - Advanced against Intermediate - the seat is suddenly worth a lot: 97% with the first move, 72% without it. Imperfect but competent play makes the tempo count; more search makes it stop counting, as the next section shows.
What Max does to the map
The same 81 openings, but with Grandmaster on one side of the board and the Strong engine line on the other - gold is Max, green is Strong:
Engine Moving first, Max wins 77 of 81 openings - Strong finds a refutation in just four. Moving second, Max refutes 64 of 81: the same second seat that won 29 games under Strong-vs-Strong wins 64 once the search deepens. Put together: under equal Strong play the first seat led 51 : 29, but give either seat to Max and it takes the game 141 : 21 overall. At engine level, search quality does not just beat tempo - it makes tempo nearly irrelevant. Two details worth keeping: bookless Max-vs-Max opens on the same column the opening book chose (one right of centre) and grinds through a 59-ply first-player win - a proper middlegame, against the book cache's 7-ply hole; and the entire Grandmaster side of this study ran on a research-only port that was verified move for move (59/59 plies) against the app engine before anything here was published.
The parity echo: where winning discs land
Classic Connect Four theory says the first player should win on odd rows and the second player on even rows (on an even-height board the second player can claim the even squares by answering on top). 9x8 has never had this measured. Across all 1,371 decisive games of this study:
Engine First-player wins finished on an odd row 54.1% of the time (737 wins); second-player wins finished on an even row 56.5% of the time (634 wins) - about ten percentage points of separation between the roles, in exactly the direction the theory predicts. A visible echo of odd/even threat structure in imperfect play, not a proof of it. Most wins at every strength complete on rows 3 to 5, where the line-participation counts peak.
What the value head believes, ply by ply
Engine Averaged over the deterministic Strong-vs-Strong opening-map games, and split by whose turn it is - because the network is a tempo optimist: at any depth it thinks the side about to move stands somewhat better, so mixing the two views into one line produces a sawtooth. Read separately, the story is clean. Through the first player's eyes (blue) the position starts moderately good - partly because the forced openings include bad replies - decays steadily, and crosses dead even around ply 35. The second-player view (gold) is the mirror shifted down. Both drift negative in long games: in grinding endgames the evaluator likes the second player, which rhymes both with the follow-up structure of an even-height board and with the second-player lean in the solved-boards table.
Data and reproducibility
The raw data is free to use (CC BY 4.0): every game with players, forced opening, winner, length, winning row and the full move list as a column string, plus the aggregates behind every chart on this page.
- line4-9x8-selfplay-v1.json - full records incl. per-ply value traces for the 82 deterministic Strong games (789 KB)
- line4-9x8-selfplay-v1-games.csv - one row per game (119 KB)
Reproducing it needs no access to anything private: the games replay from the seeds with the same engine that runs in the browser on this site (the Grandmaster level runs in the study harness only). If you compute something further with this data - or produce an exact result about 9x8 - we would love to hear about it and will cite it here: selectany@gmail.com.
Play the board yourself
The engine behind this study is playable: against the Champion in your browser, or with the full 500-puzzle Tactics set and the two Max levels in the free Android app.
