I Spent $1000 Making LLMs Play Competitive Pokémon | Doubleword

An author hosted a round-robin Pokémon tournament featuring 10 large language models playing Random Battles via the Pokémon Showdown engine, costing about $1,000 in API-equivalent spend. OpenAI's GPT-6 Astra High won the tournament, while open-weight models like DeepSeek and Kimi performed competitively in the middle tier. An analysis of decisions revealed that models were more accurate at predicting opponents from their own provider and that weaker models sacrificed Pokémon much more frequently than stronger ones. Open article →