Google’s powerful AI model, Gemini 2.5 Pro, has finished playing the old Game Boy game Pokémon Blue. Google CEO Sundar Pichai shared the news online, calling it a big win. A software engineer named Joel Z made this happen by livestreaming Gemini’s gameplay for months. Even though Joel does not work for Google, the company’s leaders cheered him on as Gemini collected gym badges and solved puzzles.
This shows how AI can learn to do tricky tasks like playing games. But Joel says it is not fair to compare Gemini with other AI models like Claude, because they use different tools.
Why Pokémon Blue Is a Big Test for AI
Pokémon Blue came out in 1996. Police force members need to develop their creatures then battle bosses as they solve advanced challenges to succeed. This makes it a good test for AI. Earlier this year, a rival AI named Claude from Anthropic started playing Pokémon Red, a similar game. Claude did well but has not finished yet. Joel Z got the idea to try Gemini after seeing Claude’s progress.
How Gemini Played the Game
Joel Z used special tools to help Gemini understand the game. Here is how it worked
Step 1
Gemini saw the game through screenshots. A grid and a small map helped it know where to go.
Step 2
Smaller AI helpers, like a “Pathfinder” and a “Boulder Puzzle Solver,” handled tough parts. These helpers were all versions of Gemini.
Step 3
Joel Z fixed some mistakes but did not give direct hints. For example, he told Gemini to talk to a bad guy twice to get a key, which was a bug in the game.
Tools Used
- An emulator to run the game
- Screenshots with maps and grids
- Memory summaries to stay on track
Is Gemini Better Than Claude?
Anthropic’s Claude AI is still playing Pokémon Red. It gets stuck sometimes and has to restart. Gemini finished Blue faster, but Joel Z says they cannot be compared fairly. Claude uses different tools and gets different information. Both AIs need lots of help to play.
What This Means for AI and Gamers
The exciting victory of Gemini exists while the system still needs human assistance as well as particular hardware tools. The AI still needs human help and special tools. The appropriate setup enables AI systems tomaster complex game systems.
While watching AI play games live brings happiness to gamers who share nostalgic game memories. Through his Twitch stream Joel Z provides viewers the opportunity to observe Gemini making decisions while he plays. The system will get updates but he refuses letting viewers share strategies with Gemini.
What’s next?
AI achieved a minimal accomplishment when it completed Pokémon Blue. The accomplishment demonstrates how models similar to Gemini operate effectively in problem-solving situations requiring planning capabilities. Real independence remains unattainable in the present state of affairs. The path to innovation relies on combining AI power with human creativity at this current stage.