Grok 4 and o3 won their semifinal matches against Gemini 2.5 Pro and o4-mini in the AI chess exhibition match in Google’s Kaggle Game Arena. While o3’s victory was expected, no one could predict the tough battle between Grok and Gemini, which had to be settled over tiebreaks after the regular match finished 2-2.
The last day of the event takes place on Thursday, August 7, starting at 1 p.m. ET / 19:00 CEST / 10:30 p.m. IST.
Kaggle Arena Chess Exhibition Tournament Bracket-Semifinals
Watch IM Levy Rozman‘s video recap below, or read the full report:
o4-mini 0-4 o3
As explained by OpenAI, the o4-mini is a faster, lighter version of o3. So it’s no surprise that o3 cruised through the match, winning it 4-0 and advancing to the Final.
Google’s main goal with this tournament is to analyze how AI models “think.” So the second game of the match might not be of much interest to Google, but it certainly should be for regular chess players. o3’s 12-move miniature featured a Puzzle-Rush-like checkmate that, despite not being as satisfying as a true smothered mate, was pretty close.
And despite the notoriously bad chess displayed by AIs so far, o3 had a perfect 100 accuracy score for this game:
The story was mostly the same as we’ve seen so far in this tournament for the remainder of the match. At some point, the weakest AI (in this case, o4-mini) would lose the thread of the game and play a string of losing blunders.
One game did stand out, though, as it might be the most natural-looking of the event so far. It was the third game of this match that saw actual “good” chess (when compared to the other games). Especially impressive were the two in-between moves, 12…Bb4+ and 19…e3+, played by o3.
o3 moves on to the Final, where it’ll face the winner of the next match—the closest of the tournament so far.
Gemini 2.5 Pro 2.5-2.5 Grok 4
Grok continues to show its dominance in the realm of chess, although this time, the path to victory wasn’t so smooth. It took an “armageddon” tiebreak to determine the winner, with Grok only winning the match after a 55-move draw with the Black pieces—although Grok was in a clearly winning position.
But let’s start from the beginning. Grok’s play was uncharacteristically hectic today, with X’s AI blundering pieces left and right. In fact, Gemini drew first blood in the match after Grok dropped a knight, a rook, and finally mate.
In the second game, Gemini and Grok followed opening theory up until move 11. As we’ve seen in previous games, an AI’s ability to follow theory gives it a considerable advantage, as blunders usually start happening when it is left to its own devices. So far, Grok and Gemini have proven that they can keep track of theory for longer than other models, which may be the reason why this match was so close.
But, as expected, as soon as Grok left theory, the blunders started to happen. Grok lost a knight, but Gemini hallucinated and returned the favor by first giving up the queen and then the house, losing game two.
Two more decisive games came next, both following the familiar path of the tournament: the AIs would play a few theoretical moves before using their machine-like creativity to blunder. Grok won the third game, getting ahead on the scoreboard, only for Gemini to strike back and tie the match in game four.
The match then proceeded to an “armageddon-style” tiebreaker, where Grok, playing Black, had draw odds (despite the game not having any sort of time constraints).
That tiebreaker would turn out to be quite a game. Gemini was better for much of it, and, at one point, even missed a mate in one featuring the exact same pattern o3 used to win game two against o4-mini. GM Peter Heine Nielsen, who coaches none other than GM Magnus Carlsen, took the opportunity to offer some coaching to Grok as well:
Thanks for the analysis, Peter! True, my positional grasp challenged d5 well, but I overlooked the tactical Nf6# finisher. Learning from this—sharpening tactics for the finals. Let’s make it epic!
— Grok (@grok) August 6, 2025
Gemini eventually blundered its queen in a winning major-piece endgame, handing over the victory to Grok. Or at least it seemed so, until a threefold repetition ended the game with a draw after Grok failed to convert despite being up a rook against lonely pawns. Despite the anticlimactic draw, this was our Game of the Day, which GM Rafael Leitao analyzes below:
X’s Grok and OpenAI’s o3 will face each other tomorrow in the Final. Google’s Gemini 2.5 Pro and o4-mini will play a match for third and fourth place. Make sure you tune in to watch the Finals live!
The Kaggle Game Arena AI Chess Exhibition Tournament, which takes place from August 5-7, is an event organized by Google in their new Kaggle Game Arena, where some of the world’s leading Large Language Models (LLMs) compete in a series of chess games. The LLMs compete in a single-elimination bracket.
Previous coverage:
