Why the Best AI Systems Are Still So Bad at Pokémon (time.com)

Three of the most advanced AI systems—GPT 5.2, Claude Opus 4.5, and Gemini 3 Pro—are currently engaged in a livestream challenge to master classic Pokémon games on Twitch, highlighting their ongoing struggles despite considerable advancements. These models, while built on robust datasets, exhibit weaknesses like overconfidence and confusion, often taking non-linear paths or getting stuck in the game. For instance, Claude Opus 4.5 has spent days in-game without progressing due to issues like not recognizing tasks, while some earlier models could not play effectively at all. This experiment serves as a unique lens to evaluate AI capabilities beyond standard benchmarks, giving viewers insight into the models' real-world performance. The significance of this endeavor lies in its exploration of AI's ability to manage long-term planning and execution—skills crucial for automating cognitive work but currently lacking in these general-purpose systems. Unlike AI specifically designed for strategic games like chess, the real-time strategy required in Pokémon tests their capabilities in a different way. The results reveal a gap between knowledge and execution, with performances showcasing human-like quirks and reasoning breakdowns under pressure. Early indicators show progress, as models like Gemini 3 Pro display improved continuity in gameplay and decision-making, raising expectations for future applications of AI in complex tasks beyond gaming.