Pokémon TCG: AI Challenges Chance with $300,000 at Stake

Pokémon TCG: AI Challenges Chance with $300,000 at Stake

The Pokemon Company launched the “Pokémon TCG AI Battle Challenge” on Tuesday: to create AIs capable of playing the Standard format of the card game. $300,000 in prizes, with the finals in September in Tokyo. Hidden information and chance are the real challenges.

Since Deep Blue in 1997 and AlphaGo in 2016, major milestones in AI are often measured by their ability to master demanding human games. Chess, Go, poker (Libratus in 2017, Pluribus in 2019), StarCraft II (AlphaStar in 2019): each challenge reveals something specific about what algorithms can or cannot yet do. The Pokemon Company, Matsuo Institute, and HEROZ launched the “Pokémon TCG AI Battle Challenge” on June 16, with support from Google, Google Cloud, and Nvidia, on the Kaggle platform. The competition is open until mid-August for the simulation phase, with an in-person final in Tokyo in September.

Why Pokémon TCG is Harder than Chess for an AI

The real challenge is not where the press release suggests. Chess and Go are perfect information games: both players see the entire board at all times, with no randomness involved. Deep Blue defeated Kasparov through brute-force calculation, while AlphaGo utilized neural networks trained on tens of millions of human positions. The Pokémon TCG structurally resembles poker more: hidden hands, random draws, decisions made with incomplete information about the opponent's deck. This is why programs capable of defeating the best human poker players only emerged between 2017 and 2019 (Libratus and then Pluribus, from Carnegie Mellon and Meta), twenty years after the first dominance in chess.

AI has defeated Kasparov, mastered Go, and held its own against the best poker players. Now it will have to manage a Dragonite that has just been played.

Specifically for this challenge: the Standard format offers a pool of 2,000 cards, of which 60 make up a deck. The cards are not visible to the opponent, draws are random, and abilities interact in a combinatorial manner. The AI will need to handle uncertainty with each move, not just calculate variations on a static board. A useful clarification to avoid confusion with current industry trends: the competition targets traditional machine learning (reinforcement learning, rule-based systems), the same type of algorithm used in chess programs for decades, not generative AI.

A First for The Pokemon Company, Already Dividing the Community

Claude 3.7 Sonnet faced off against Pokémon Red on Twitch in early 2025, in a demonstration organized by Anthropic to showcase the model's agentic capabilities. Gemini 2.5 Pro followed a few months later, with a curiosity noted by DeepMind: observable “panic” phases in the model's behavior during tense battles, characterized by a degradation of reasoning under pressure. These experiences focused on single-player open-world video games; the challenge launched this week tackles competitive card games, a structurally different problem.

AI has defeated Kasparov, mastered Go, and held its own against the best poker players. Now it will have to manage a Dragonite that has just been played.

A few clarifications regarding the “you can participate” in the title: the competition is open to developers worldwide to submit AI agents via Kaggle, not to the general public for play. The Strategy category additionally requires a report explaining the logic behind the agent. The top eight teams will advance to the finals in Tokyo, with $50,000 for first place and $30,000 for second.

One governance detail to note: this is the first time The Pokemon Company has publicly taken a stance in support of AI (even if we are talking about traditional algorithms rather than generative AI), which contrasts with the position of Nintendo, one of its three co-owners (along with Game Freak and Creatures), which categorically refuses to integrate generative AI into its own games.