Google’s Gemini 2.5 Pro Conquers Pokémon Blue

Google’s most advanced AI model, Gemini 2.5 Pro, has achieved a significant milestone by completing the 29-year-old video game Pokémon Blue. The achievement, which signifies a leap in how AI models handle complex, multi-step tasks, was publicly celebrated by Google CEO Sundar Pichai.

The feat was documented through a Gemini Plays Pokemon livestream, a project spearheaded by independent software engineer Joel Z. While Google leadership, including product lead for Google AI Studio Logan Kilpatrick, closely monitored the progress, the development is credited to an independent developer rather than a direct Google initiative.

The Role of Agent Harnesses in AI Gaming

The ability of an LLM to play a classic title like Pokémon is not a native skill but the result of “agent harnesses.” These frameworks provide the model with:

• Real-time screenshots of the game.
• Additional contextual data overlays.
• The ability to trigger specialized agents to make decisions.
• Mechanisms to execute button inputs based on the AI’s reasoning.

Before this success, Google executives had been tracking the model’s performance, with Kilpatrick noting last month that Gemini had secured its fifth gym badge, outpacing other models. This performance sparked a lighthearted comment from Pichai, who joked that the company was developing “Artificial Pokémon Intelligence.”

Contextualizing AI Benchmarks

The pursuit of gaming milestones gained traction in February when Anthropic showcased its Claude models playing “Pokémon Red.” Anthropic credited the success to “extended thinking and agent training,” which allow the AI to navigate unexpected scenarios. Despite the rivalry, Joel Z has cautioned against using these sessions as a direct performance benchmark.

“Please don’t consider this a benchmark for how well an LLM can play Pokemon,” Joel Z stated on his Twitch channel. He emphasized that Gemini and Claude operate under different parameters, utilizing unique tools and information streams that make direct comparisons difficult.

Addressing the Question of “Cheating”

While the AI required external support to finish the game, Joel Z maintains that the interventions used were designed to enhance decision-making rather than bypass challenges. According to the developer, the framework avoids providing walkthroughs or direct instructions for specific puzzles.

The only exception mentioned involved a specific workaround for a known bug: informing the model that it needed to speak to a Rocket Grunt twice to obtain the Lift Key—a requirement that was later adjusted in the subsequent release, Pokémon Yellow. As development continues, Joel Z notes that the framework remains in an evolutionary state, with further refinements expected for the project.