OpenAI’s o3 model has been tasked with playing Pokemon Red and achieving the ultimate goal of defeating the Elite Four to become Kanto’s champion.
The experiment, being streamed on Twitch since May 27, has the AI provide its reasoning for every individual decision displayed on-screen as it attempts to complete a series of objectives. While the overall aim is for o3 to complete the entire game, that goal is divided into smaller directives.
As of writing, the model has managed to obtain two of the eight Gym Badges required to access the endgame Elite Four and is now on its way to Vermilion City to hitch a ride on the S.S. Anne.
How long that will take remains to be seen, however. Every action, including movement and battles, is assessed for a significant amount of time to find the most efficient action.
What is OpenAI o3?
Unlike ChatGPT, which is a general-purpose chatbot, OpenAI o3 is essentially a more specialized model designed to have superior reasoning for problem-solving.
Why Pokemon Red was chosen to showcase these qualities, however, isn’t clear. The Game Boy classic’s slow-paced gameplay, married with several semi-complex mechanics, is one possible answer. Whatever the reason, this is far from the first time Pokemon Red and Blue have been used for experimental endeavours.
In 2014, TwitchPlaysPokemon allowed viewers to collectively play Pokemon Red by voting on inputs for the game to make remotely. The stream eventually went viral and ended with the defeat of final boss, Blue, after 16 days.
This isn’t the first instance Pokemon being used to put AI through its paces, either. In April 2025, a software engineer programmed Google’s Gemini to play Pokemon Blue.
