Google’s Gemini AI Completes Pokémon Blue with Developer Assistance and Custom Tools

Google’s Gemini 2.5 Pro, the company’s most advanced AI model, has achieved a notable milestone by completing the classic 1996 Game Boy game Pokémon Blue. The accomplishment was celebrated by Google CEO Sundar Pichai, but the feat was actually orchestrated by Joel Z, an independent software engineer unaffiliated with Google. The project, broadcast on Twitch, involved significant developer interventions and a sophisticated agent harness setup. Gemini played the game by analyzing screenshots and game data, making decisions, and issuing commands, sometimes delegating complex tasks to specialized sub-agents also powered by Gemini. To support the AI, the developer provided grid overlays, memory management, and occasional clarifications, such as addressing a known game bug. However, no direct walkthroughs or explicit instructions for specific challenges were given. The project highlights the growing use of classic games like Pokémon as benchmarks for evaluating AI reasoning and planning abilities. While Anthropic’s Claude model has made progress in Pokémon Red, it has not yet completed the game. Joel Z and other experts caution against using these results as direct benchmarks for AI comparison, as each system operates under different conditions and receives different types of assistance. The Gemini Plays Pokémon framework is still evolving, with ongoing improvements aimed at reducing developer intervention and enhancing the AI’s autonomy. This experiment demonstrates both the promise and current limitations of large language models in complex, open-ended environments.