RAPFI Wins Five of Six Divisions at Gomocup 2026
At the 2026 edition of gomoku's annual machine tournament, one program won five of six rule-variant divisions outright — and the one it lost went to an entirely different podium.
Gomocup, the annual world championship for computer programs playing gomoku (five-in-a-row), ran its 2026 edition on June 5-7, an online, submission-based event capped at 1GB of memory and 256MB per program. When the results were tallied, one engine, RAPFI, had won five of the tournament's six divisions outright: Freestyle on both the 20x20 and 15x15 boards, the speed format Fastgame, Renju, and Standard. The sixth division, Caro, went to a different program entirely.
The Elo ratings published afterward show how wide the gap is. In the Freestyle 20x20 ranking, RAPFI sits at 3073, 194 points above the second-placed program, Katagomo 2026, at 2879. In Freestyle 15x15 the margin narrowed to 9 Elo, but RAPFI still finished ahead. It also leads Fastgame at 2932 and Renju at 3202, and topped Standard as well, with Katagomo second, AlphaGomoku Kozarzewski third, and a new program, Chloris, fourth.
The six divisions are not the same game played six times. Freestyle allows a win with five or more stones in a row, on either a 20x20 or 15x15 board; Standard, also 15x15, requires exactly five; Renju forbids the first player from making certain double-threat patterns that would otherwise make moving first overwhelming; Caro requires the five-stone line to be unblocked on both ends, which changes which shapes count as a lethal threat and which lines are worth calculating. A program's search and evaluation logic has to be tuned separately for each of these, which is why a single engine dominating all of them is not automatic.
In Caro, the winner was AlphaGomoku, an engine written by Maciej Kozarzewski, with JAX second and Embryo third — none of them the program that topped the other five boards. RAPFI did not place on the Caro podium at all. Gomocup's results page states the standings without offering a reason for the gap, and none is evident from the numbers alone; what the results show is that the ruleset produces a genuinely different competitive order, not RAPFI merely slipping to second or third under conditions it wins elsewhere.
The tournament's other notable result belonged to a program with no prior Gomocup history. Chloris, submitted by Nelly Semenova, went undefeated through its qualifying league in the Standard division before advancing to finish fourth overall, with a Standard Elo of 2607 against RAPFI's 2791.
Gomocup has run every year since 2000, making it one of the longest-continuously-running machine-versus-machine tournaments of any kind, predating both the chess-engine circuits and the LLM game arenas that draw more attention today. Its scoring has used a Bayesian Elo system since 2017, the same family of method used to rate engines on TCEC, which is what makes a claim like a 194-point gap comparable across seasons rather than just a single-event score.
The next edition's submission deadline typically falls in late May. As of this writing, Gomocup's organizers, Kai Sun and Tianyi Hao, have not announced changes to the ruleset or division structure for 2027.