Pokémon's TCG AI Contest Splits Ranking From Prize Money

Submissions closed on August 16, 2026 for the automated ladder at the center of the Pokémon Trading Card Game AI Battle Challenge, a competition where software agents play the physical card game against each other rather than against humans. The ladder does not stop there: games keep running for roughly two more weeks so ratings can settle before the standings lock, which puts the field in its final stretch as of this writing.

The contest is organized by The Pokémon Company together with Kaggle, the Google-owned competition platform, the University of Tokyo's Matsuo Institute, and the Japanese AI firm HEROZ, with additional support from Google Cloud and Nvidia. It was announced on June 16, 2026, carries a combined prize pool above $300000, and is scheduled to conclude with an in-person final in Tokyo, streamed on Pokémon's official YouTube channel.

The Simulation Category works like most machine-learning ladders: teams submit up to five agents a day, and only the two most recent submissions count toward a team's rating, an automated Elo-style score that rises and falls with each result. Agents play out full matches against opponents of similar rating, unattended, which is what produces the roughly two-week settling period after submissions close — a rating needs enough games against enough opponents before it means anything.

What sets the format apart is where the money sits. The Simulation ladder itself pays nothing; the organizers describe it as a proving ground. Prize money is attached to a separate Strategy Category, open only to teams that fielded an agent in Simulation, which requires a written report of up to 2000 words explaining the agent's approach. Judging weights the modeling approach at 70 percent, the deck concept at 20 percent, and the quality of the report itself at 10 percent, due September 14, 2026.

Eight Strategy Category teams will be selected to advance, each receiving $30000 regardless of how the live final goes. At the Tokyo final, the winning team takes $50000 and the runner-up $30000; every team that reaches the final also receives $3000 in Google Cloud credit.

Separating ranking from reward is not how most agent-vs-agent contests are built — Kaggle's own Game Arena and general leaderboards like Arena pay entirely in standings, not cash, and rank purely on win rate. Requiring a judged report on top of a ladder result is a deliberate hedge against agents tuned only to farm rating against a narrow set of ladder opponents, at the cost of asking every serious team to also write up, and defend, why their agent plays the way it does.

The ladder's final standings are expected within days; where the Tokyo final falls on the calendar has not been fixed in the public materials reviewed for this article.