Less than a week after OpenAI introduced its latest model, GPT-6 Astra, a user ran a test in which the model autonomously completed Valve's first-person puzzle-platform game Portal. The experiment was carried out by a user named CozyBlade, according to Tom's Hardware.
Tools and method
CozyBlade used the Codex Pro subscription, OpenAI's software development agent offering, to enable GPT-6 Astra to control the game. During the run the AI received screenshots and data about the player character's position.
How the run proceeded
Gameplay paused while the model "thought"—that is, while it processed inputs and determined the next action. When the AI announced what to try next, the game resumed and executed the suggested move. The entire process, including pauses for the model's reasoning and the in-game actions, took about 24 hours to complete the game.
A two-hour video compilation was produced from the run, which the author says consists entirely of AI-driven gameplay.
Further details
Implementation details for CozyBlade's solution are available on GitHub for those who want to examine how the setup was constructed and how the interaction between the game and the model was managed.
Why this matters
The demonstration highlights both the practical capabilities of GPT-6 Astra in extended interactive tasks and the concerns voiced by some researchers about powerful models. In the short time since GPT-6 Astra's unveiling, the Portal run provides a concrete example of how such models can perform automated, long-duration decision-making in complex environments.
