Claude struggles to progress in Pokémon Red game

Anthropic's AI agent, Claude, is currently attempting to beat the classic game Pokémon Red on a livestream hosted on Twitch. This unique gaming project began approximately a month ago, but Claude has not made significant progress so far. Initially, earlier versions of Claude struggled with simple tasks in the game. For instance, the Claude 3.5 version would often run away during battles. In contrast, Claude 3.7 Sonnet, released in February 2025, showed improvement. It quickly defeated game characters Brock and Misty, showcasing better planning and learning abilities. Despite this progress, Claude 3.7's advancement in the game seems to have slowed down. It took the AI 78 hours to navigate through Mt. Moon, a section that typically takes a human player only a few hours. Viewers have noticed that Claude often moves in circles and runs into walls. The livestream remains interesting as it displays Claude's thought process while making decisions. While the AI performs well in text-based challenges, it struggles with navigation in the game's visual environment. Overall, Claude 3.7 has made advancements compared to its predecessors, but it still has a long way to go, with 151 Pokémon left to catch.