Compared with chess or Go, Carcassonne has barely been studied. For this page, the strongest AI on this site played 2,000 games against itself, and we measured the first-player advantage, where points come from, how often games turn around, what each tile is worth, and more. We also compare Carcassonne with chess, Go, shogi and other games.
Over 2,000 games: first player 1,123 wins, second player 838 wins, 39 draws. The first player scored 57.1% (95% interval 54.9%–59.3%) and won by +4.0 points on average (±0.5). The first player has a clear edge (very unlikely to be chance).
Wins for the first and second player (2,000 games)
The main reason is the tile count. The 71 tiles after the start tile are placed in turn, so the first player places 36 and the second player 35. The first player also places the very last tile. Removing one tile so both place the same number almost erased the gap (section 9). Draws are rare: 1.9%.
The average score per player was 98. Winners averaged 106 and losers 90. The median margin was 14 points; 38% of games were decided by 10 points or less, and 22% by 5 or less.
Winning margins (5-point bins)
Even between strong AIs, more than one game in three ends within 10 points. A few points from a short road, or 3 points from a field at the end, decide the game.
Cities are the biggest source at 43% of all points; fields (farmers) gave 17%. What people often miss: points scored at the end of the game (unfinished cities, roads and monasteries, plus fields) make up 43% of the total.
Points per player per game, by source
Of the features that had a meeple, this is the share still unfinished when the game ended. 50% of monasteries end unfinished (cities 30%, roads 18%).
Share of occupied features left unfinished
Average banked score (completed features) by turn. Scoring speeds up later as big cities get completed, and jumps at the very end when the final scoring is added.
Average score by turn (first player, second player)
The player leading on projected points (the score if the game ended now) went on to win 60% of the time in the opening (turn 17), 67% at halftime (turn 35), and 79% late in the game (turn 52). The lead changes hands 5 times per game on average. A player 15 or more points behind at halftime still came back to win in 18% of those games.
Projected lead at halftime vs. how often the leader wins
An early lead means little. Even if you are behind in the middle game, you can catch up with features that are about to close and with fields. Compare projected points, not banked points.
The strong AI places a meeple about 16 times per player per game, on 45% of its turns. It can place more than twice its 7 meeples because it keeps completing features and getting them back. Even so, it had none in hand on 13% of its turns.
Average meeples in hand (first player, second player)
Where meeples were placed
Placements: cities 43%, roads 32%, monasteries 14%, fields 11%. At the end, each player still had 4.1 meeples on unfinished features on average, and 1.6 farmers that scored.
Keep meeples cycling: place, complete, place again. Use small cities and short roads that close quickly, and try not to run out before the endgame.
When a player draws a tile, how much the projected score difference moves in their favour on that turn, on average. The best tiles were Monastery with road, Monastery, Two adjacent city edges; the least valuable were Four-sided city (shield), City tube, Three-sided city (the × number is how many are in the game).
| Tile | Gain for the drawer (pts) |
|---|---|
| Monastery with road×2 | +4.8 |
| Monastery×4 | +4.5 |
| Two adjacent city edges×2 | +3.4 |
| City edge + curve (left)×3 | +3.2 |
| City edge + curve (right)×3 | +3.1 |
| Two opposite city edges×3 | +3.1 |
| City edge + straight road×4 | +3.1 |
| One city edge×5 | +3.0 |
| City edge + junction×3 | +3.0 |
| Corner city (shield)×2 | +2.5 |
| Corner city + curved road (shield)×2 | +2.5 |
| Three-sided city + road (shield)×2 | +2.4 |
| City tube (shield)×2 | +2.3 |
| Three-sided city (shield)×1 | +2.2 |
| Junction×4 | +1.9 |
| Three-sided city + road×1 | +1.9 |
| Crossroads×1 | +1.9 |
| Curved road×9 | +1.8 |
| Corner city + curved road×3 | +1.8 |
| Corner city×3 | +1.7 |
| Straight road×8 | +1.6 |
| Four-sided city (shield)×1 | +1.5 |
| City tube×1 | +1.4 |
| Three-sided city×3 | +1.1 |
A monastery tile immediately counts the tiles around it, so drawing one is a big chance. Common road tiles are worth little on their own, but how you use them still matters, for example by not filling the spaces around your opponent's cities and monasteries.
The luck in Carcassonne is which tile you draw and when. We played every tile order twice and compared. With the same tiles, the same player won only 55% of the time (100% if the tiles decided everything, 50% if they did not matter at all). About 17% of the spread in final margins can be put down to the tile order.
We also played AIs of different strength against each other, two games per tile order with seats swapped. The last column converts each score into the rating gap (Elo) it would mean in chess.
| Stronger | Weaker | Games | Stronger scores | Upsets | Avg margin | Chess Elo gap |
|---|---|---|---|---|---|---|
| Greedy AI | Random AI | 400 | 100.0% | 0.0% | +71.6 | — |
| “Standard” on the play page | Greedy AI | 400 | 92.4% | 7.2% | +28.9 | +433 |
| “Strongest” on the play page | “Standard” on the play page | 120 | 69.6% | 28.3% | +10.2 | +144 |
The play page's “Standard” AI scores 92% against an AI that only chases immediate points; the weaker side wins just 7% of games. Single games swing with the draw, but over many games the better player comes out ahead. It is worth reviewing your play instead of blaming the tiles.
A Carcassonne turn offers about 81 moves on average (tile spot and rotation × meeple choice), and up to 444. Counting only the tile spot and rotation, it is 23. The game tree (all possible move sequences in a game) is about 10127, similar to chess. Adding the branching from the tile draw (about 17 kinds of tile) gives about 10209.
| Game | First-move advantage | Luck | Choices per move | Game-tree size | AI milestone |
|---|---|---|---|---|---|
| Carcassonne | First player scores 57.1% (measured here) | Order of the tiles drawn (public) | about 81 (measured) | 10127 (measured, draws excluded) | No public record of beating top humans yet |
| Chess | White scores about 55% (W 37.5%, D 34.9%, L 27.6%) [*] | None | 35 | 10123 | 1997: Deep Blue beat world champion Kasparov 3½–2½ [*] |
| Go | Black moves first, so White gets 6.5 points of komi (Japan). With 5.5 komi Black won 51.855% [*] | None | 250 | 10505 | 2016: AlphaGo beat Lee Sedol 4–1 [*] |
| Shogi | — | None | 92 | 10226 | 2017: Ponanza beat Meijin Amahiko Sato 2–0 [*] |
| Othello | A draw with perfect play (solved in 2023) [*] | None | 10 | 1058 | 2023: solved (a draw) [*] |
| Checkers | A draw with perfect play (solved in 2007) [*] | None | 2.8 | 1040 | 2007: solved (a draw) [*] |
| Backgammon | — | Dice | 250 | 10144 | 1990s: TD-Gammon taught itself with a neural net [*] |
| Poker (heads-up) | — | Cards dealt (hidden hands) | — | — | 2017: Libratus beat four top pros [*] |
[*] links to the source. Choices per move and game-tree size for other games come from the table in Wikipedia's “Game complexity” (rounded-up powers of ten). Each game is counted differently, so compare only the order of magnitude. The chess first-move figure is from the chessgames.com database (2015).
In Go, White receives “komi”, extra points that offset Black’s advantage of moving first. To see whether the same works in Carcassonne, we changed the rules and had the AI play at least 1,000 games under each.
| Rule | Games | First player scores | 95% interval | If ties go to the second player | First player's avg margin |
|---|---|---|---|---|---|
| Current rules (71 tiles, no adjustment) | 2,000 | 57.1% | 54.9%–59.3% | 56.1% | +4.0 |
| Second player starts with 4 points (komi 4) | 1,000 | 52.8% | 49.8%–55.9% | 51.7% | +1.3 |
| Remove 1 tile face-down (35 tiles each) | 1,000 | 51.2% | 48.1%–54.3% | 50.4% | +0.5 |
With “Second player starts with 4 points (komi 4)”, the first player scored 52.8% (average margin +1.3); the gap is within the margin of error. With “Remove 1 tile face-down (35 tiles each)”, the first player scored 51.2% (average margin +0.5); the gap is within the margin of error.
Under the standard rules, tied players share the victory. But at the Japanese Championship and the German championship, a tie is a win for the second player. The “If ties go to the second player” column in the table counts games this way. Under the current rules the first player then wins 56.1% (or scores 57.1% when a draw counts as half a win). Because draws are rare to begin with, this lowers the first player's rate by only 1.0 points, so almost all of the first-player edge remains. Our estimate: if ties go to the second player and the second player also starts with 3 points, the first player wins 51.0%; with 4 points, 48.8%, roughly even.
The chart below estimates what happens if the second player’s bonus is added to the margins of games played under the current rules. 3 points is not enough, and 5 points tips the balance to the second player.
First player’s score for each starting bonus given to the second player (estimate)
When only a few tiles are left in the bag, we can check every possible draw and every move by both players, and find exactly the move with the highest chance of winning. This is the game-theoretic best move, the idea poker players call “GTO”. In Carcassonne both players can see the board and which tiles (and how many) are left; the only hidden thing is the luck of the draw order. So unlike poker there is no need to mix moves at random: the best move is always a definite move.
We solved the last 4 moves (3 to 0 tiles left in the bag; 1,600 positions) of 400 AI-vs-AI games.
| Tiles left in the bag | Positions | Positions where the move matters | AI found the best move |
|---|---|---|---|
| 3 left | 400 | 101 (25%) | 76% |
| 2 left | 400 | 67 (17%) | 84% |
| 1 left | 400 | 38 (10%) | 100% |
| 0 left (last tile) | 400 | 37 (9%) | 100% |
In 85% of the last-4-move positions, the chance of winning was the same wherever you played. Positions where the move changes the outcome made up 20% when the score gap before the move was 20 points or less, and 5% when it was larger. The same search as “Strongest” on the play page (gen1144) found the best move in 86% of the positions where the move matters. It only missed with 2 or 3 tiles left, worth about 1.4 percentage points of win rate per player. The play page AI now uses this exact solving once 3 tiles are left in the bag.
The move that scores the most points and the move with the best chance of winning differed in only 5 of 1,200 positions (checked with 2 or fewer tiles left).
The solver searches to the end of the game, weighting each kind of remaining tile by its draw probability and assuming the player to move picks the move with the highest chance of winning (a draw counts as half a win). This is the expectiminimax method (Russell & Norvig, AIMA chapter 6). With 3 tiles left it usually takes a few seconds or less per move (the play page AI stops after 3 seconds and plays the best move it has verified so far). With 4 tiles a single position can take minutes, so the play page does not use it there. Positions come from AI-vs-AI game records, and the “Strongest” move was found by searching each position again. Code: src/carcassonne/endgame.py and scripts/gto_endgame.py; results: docs/results/gto_endgame.json.
The AI uses the settings of “Standard” on the play page at the time of collection (2000 searches per move with the evaluation function gen975; since 2026-11-11 “Standard” also re-checks its top candidates against the opponent’s best reply and is a bit stronger). Each of 1,000 tile orders was played twice with different move randomness (as of 2026-11-05). Score intervals are Wilson 95% confidence intervals; the first-player test is a binomial test excluding draws (p < 0.001). “Projected points” count unfinished features and fields as if the game ended at that moment. Matches between different AIs pair two games per tile order with seats swapped. In the rule-change games the AI knows the rule (komi is the second player's starting score; with 70 tiles the last tile stays face-down and unused). Numbers for human games may differ. Scripts: scripts/collect_game_stats.py and scripts/summarize_game_stats.py; results: docs/results/game_stats.json.