- Google’s Gemini 2.5 Pro has formally completed Pokémon Blue
- The game went as a direct experience by an independent engineer
- Gemini played the game with some mild developer intervention, but mostly himself
Google’s Gemini Ai may not have passed the touring test yet, but it will be very famous in the courtyard of the school after winning the Pokémon Blue game three decades ago. Gemini 2.5 Pro is now Google’s latest AI model and a Pokémon Master, as shown in a Tweech Live Stream, named “Gemini Play Pox Pokémon” run by an engineer named Google.
What end! Gemini 2.5 Pro just completed Pokémon Blue! Special thanks to @Kodfjul for a direct series and running a series, and everyone who makes the semen on the way. pic.twitter.com/e2PN3TPFEBMay 3, 2025
You may wonder why the AI model, which defeated the 30 -year -old game, paid so much attention. It is partly due to the spectacle, but also because of the enmity of the AI model. In February, Anthropic showcased the progress of his cloud model in defeating Pokémon Red. He used the game to show the game “Extension Thinking and Agent Training” and launched an impressive launch of the “Claude Play Pox Pokémon” Tweet Stream, Joel Z.
Before Gemini was crowned as a true Ash Kyrgyz, it is worth considering some warnings. On the one hand, Claude has not yet defeated Pokémon Red technically, but it does not automatically improve Gemini, as he employed various tools, called “Agent Harren”. Models do not play directly like a human being with a controller. Instead, he is fed screenshots of the game environment with key information overlays, then asked to create the next best action. This decision is then translated into an original button press in the game.
And Gemini is not going completely alone. Joel acknowledged that he sometimes stepped in to improve, though he had just made a point to do so to improve Gemini’s reasoning. He has also planned to continue work on the Gemini Playmone project to further improve.
Pokémon Ai
What makes it more than a strange Internet stunt means where AI is headed. Playing a game like Pokémon Blue is not about memorizing fast anxiety or controller inputs. It is about long -term strategies, adapting to surprises, and navigates ambiguous challenges. These are all sectors where AI usually needs improvement. She could not only keep her job but also end the game (at least with nudge) suggests that models like this are improving in expansion strategies.
This is also a milestone that the average person can understand. What AI is doing when you wander around in Lavender Town or read the war plan wrongly, and in this context you can understand the choice you want to be intuitively. Of course, you should not be more than what it means. The AI can now end a game that you may have defeated in middle school, but it also highlights how much human efforts still try to find AI independent.
Whether Claude or Gemini become the real poke masters, it doesn’t matter what they are playing for the development of AI. To show that AI will not just crush numbers or spam emails, people think about what AI can do, even with help. And if the AI models begin to learn how to operate in an unexpected, open environment, beating Moto can be just a very deep stone. Or at least, slightly more fruitful.