Gemini Beats Pokémon Blue: Google AI's Gaming Triumph

Gemini 2.5 Pro Achieves Victory in Pokémon Blue

Google’s most advanced AI model, Gemini 2.5 Pro, has reportedly reached a significant accomplishment: successfully completing the classic video game, Pokémon Blue.

Sundar Pichai Announces the Milestone

Google CEO Sundar Pichai announced the achievement on X (formerly Twitter), expressing excitement over Gemini 2.5 Pro’s completion of Pokémon Blue.

The Role of Joel Z

The “Gemini Plays Pokemon” livestream was developed by Joel Z, a software engineer not directly employed by Google. He has been working independently on this project.

However, Google leadership has actively supported and celebrated Joel Z’s efforts.

Progress and Comparisons to Claude

Logan Kilpatrick, a product lead at Google AI Studio, previously noted Gemini’s substantial progress in Pokémon, highlighting that it had earned five badges, surpassing other models.

This led to a playful response from Pichai regarding the development of an “Artificial Pokémon Intelligence” API.

The endeavor draws inspiration from Anthropic’s Claude AI, which has also been making strides in playing Pokémon Red. Anthropic emphasized that Claude’s capabilities in extended thinking and agent training provide a considerable advantage in tackling complex tasks, such as mastering a classic game.

A dedicated Twitch channel, “Claude Plays Pokemon,” served as a source of inspiration for Joel Z.

Is Gemini Superior?

While Claude has made progress, it hasn’t yet finished Pokémon Red. However, Joel Z cautions against direct comparisons.

He emphasizes that Gemini and Claude utilize different tools and receive varying levels of information, making a straightforward benchmark impossible.

The Importance of Agent Harnesses

Both AI models require assistance in gameplay, facilitated by “agent harnesses.” These harnesses provide the models with game screenshots enhanced with supplementary data.

This allows the AI to analyze the situation, determine the appropriate action, and then execute it through corresponding button presses.

Acknowledging Developer Interventions

Joel Z admits to making “dev interventions” to aid Gemini’s progress, but maintains that these actions do not constitute cheating.

He explains that these interventions are designed to enhance Gemini’s decision-making and reasoning skills, rather than providing specific solutions.

Specific Assistance Provided

For instance, he clarified that he didn’t offer walkthroughs or direct instructions for challenging areas like Mt. Moon.

The only significant assistance involved informing Gemini about a bug requiring interaction with a Rocket Grunt twice to obtain the Lift Key, a flaw later corrected in Pokémon Yellow.

Ongoing Development

Joel Z also points out that “Gemini Plays Pokémon” is an ongoing project, with the underlying framework continuously evolving and improving.