Google's Gemini 2.5 Pro Beats Pokémon Blue, Showing AI's Gaming Skills
Google's priciest AI model to date has achieved a remarkable feat: it has conquered a video game that has been around for nearly three decades.
Just last night, triumph was in the air when Google's CEO Sundar Pichai joyfully shared that Gemini 2.5 Pro had successfully completed Pokémon Blue!
According to the record, Joel Z, 30-year-old software developers who calls himself independent, is the ardent creator of the Gemini Plays Pokemon livestream. Even so, Google's top brass have been actively rooting for his endeavor.
Logan Kilpatrick, the product head of Google AI Studio, had actually previously commended Gemini's development in Pokémon and even earned its 5th badge. Pichai made a clever remark about creating Artificial Pokémon Intelligence, or APIs for short, in response to this development.
Why the connection to Pokémon? Well, back in February, Anthropic shared the advancements its Claude AI models were making in Pokémon Red. Claude's capability to extend its thinking and train gave it suggested potential in handling other out-of-the-box tasks - like playing vintage video games. And yes, Pokémon Red and Blue are distinct versions of an early GameBoy game that launched the now-iconic Pokémon franchise back in 1996. It even paved the way for a Claude Plays Pokemon Twitch channel, which Joel Z cites as an influence.
But despite its progress, it seems Claude hasn't yet conquered Pokémon Red. Does this make Gemini the superior Pokémon player out of the two AIs? Joel Z shed some light on this on his Twitch page, urging viewers not to benchmark the two based on their Pokémon skills since both Gemini and Claude employ different tools and process distinct information.
Both AI models also need support to play - with "Agent Harnesses" stepping up to the plate. These provide the AI with game screenshots peppered with supplementary data, allowing the AI model to strategize its next steps (which might include calling on specialized agents), and then execute the AI's decision by clicking the corresponding button.
Joel Z candidly admits that Gemini needed a bit of assistance to complete the game, but vouches for its integrity. "My aid enhances Gemini's overall reasoning and decision-making capabilities," he clarifies, "but I'm careful not to tip the scale too much. Take note that there are no walkthroughs or specific guidance for any particular difficulties.
In addition, he adds, "Gemini Plays Pokémon is very much a work in progress, and the platform is continually evolving."