OpenAI’s latest experiment is pretty wild: their specialized o3 model is playing Pokémon Red live on Twitch, eager to conquer the Elite Four and become the Kanto champion. What makes this so interesting? Every move the AI makes is broadcast with its reasoning, letting us see how it thinks through the game’s challenges.
The stream started on May 27 at OpenAI’s Twitch channel. The AI has earned two of the eight Gym Badges needed to unlock the Elite Four and is heading toward Vermilion City, ready to board the S.S. Anne. But don’t expect speedruns here: the model takes its time, analyzing each step carefully to pick the most efficient action. Watching an AI decide whether to move left or right for several seconds is mesmerizing… and a little funny!
What is OpenAI o3?
Unlike ChatGPT, which chats about almost anything, the o3 model is explicitly built for problem-solving with deeper reasoning. Why pick Pokémon Red? Maybe because it’s a classic with slow-paced gameplay and some tricky mechanics. It’s not the first time this game has been used in experiments either — remember Twitch Plays Pokémon back in 2014? Thousands of players voted on moves to control the game collectively, and after 16 days, they finally defeated the final boss, Blue. Wow, talk about patience!
Pokémon games seem to be a popular testbed for AI. In April 2025, a software engineer taught Google’s Gemini AI to play Pokémon Blue. These experiments show how AI can tackle complex tasks in gaming environments, and watching O3’s live playthrough is a rare chance to see that process unfold in real time.
Curious how the AI plans and explains each move? Check out the stream and watch it analyze the map, save memories, and think out loud. It’s like having a super nerdy teammate who never stops discussing strategy! What do you think about AI playing classic games like Pokémon? Is it just a fun tech showcase, or could it hint at future gaming experiences? Let me know in the comments below!