Gemini AI Beats Classic Pokémon Blue Game

Cansın Cengiz

11 May 2025 — 2 min read

Google's advanced AI model, Gemini, has reportedly achieved a significant feat: completing the classic 1996 video game, Pokémon Blue. This development showcases the evolving capabilities of AI in complex, interactive environments.

The Independent Effort Behind the Victory

The achievement was spotlighted by Google CEO Sundar Pichai on X, celebrating Gemini 2.5 Pro's completion of the game. However, the "Gemini Plays Pokemon" livestream is an independent project by software engineer Joel Z, who is not affiliated with Google. Despite this, Google executives, including Logan Kilpatrick, Google AI Studio's product lead, have been vocally supportive of the initiative. Pichai even humorously remarked about developing an "Artificial Pokémon Intelligence" API.

A New Challenger in AI Gaming

This isn't the first time AI has ventured into the world of Pokémon. Anthropic's Claude AI models have also been making headway in "Pokémon Red," with the company highlighting its "extended thinking and agent training" capabilities. Joel Z, inspired by the "Claude Plays Pokemon" Twitch channel, emphasizes that these AI gaming endeavors shouldn't be seen as direct benchmarks. He points out that "Gemini and Claude have different tools and receive different information," making direct comparisons difficult.

How AI Navigates the Pokémon World

Both Gemini and Claude require assistance to play these games, primarily through "agent harnesses." These systems provide the AI models with game screenshots enriched with additional information. The AI then decides on the next action, which might involve calling specialized agents, and the system executes the corresponding button press in the game.

The Role of Human Intervention

Joel Z acknowledged that "dev interventions" were made to help Gemini complete Pokémon Blue. He clarified that these interventions were aimed at improving Gemini's overall decision-making and reasoning, not providing specific game solutions or walkthroughs. One notable exception was informing Gemini about a game bug (later fixed in Pokémon Yellow) requiring interaction with a character twice to obtain a key item. "My interventions improve Gemini’s overall decision-making and reasoning abilities," Joel Z stated. He also mentioned that the "Gemini Plays Pokémon" framework is continuously evolving.

Looking Ahead

Gemini's success in Pokémon Blue, even with assistance, marks an interesting moment in AI development, demonstrating progress in AI's ability to tackle long-horizon tasks and interact with complex systems. As these AI models and their "agent harnesses" continue to develop, we may see even more impressive feats in the future.