The GPT-6 Astra model, recently announced by OpenAI, has completed the puzzle game Portal without human intervention. The experiment was conducted by a user named cozyblaze, who integrated the AI agent with the game to solve puzzles within the test chambers. The system processed visual data from the game and received player coordinates to navigate the environment.
The AI operated using the new Astra context management system, treating the game as a turn-based strategy experience. The process required nearly 24 hours to complete, with the system pausing frequently to analyze screenshots and determine subsequent actions. While the final video summary lasts two hours, the model processed the entire game independently.
To successfully navigate all obstacles, the AI model performed 3,336 tool calls. The total cost for the API usage during this experiment amounted to $571.18. This achievement aligns with a long-term goal established by OpenAI in 2016, which aimed to develop a single agent capable of mastering a wide range of games.