GPT-6 Astra completed the puzzle game Portal without human intervention after the initial goal was set, according to developer cozyblaze. The run lasted about 23 hours and 43 minutes and ended at the game’s credits. Code and documentation for the experiment were published on GitHub.
The model controlled the game through the Model Context Protocol and a modified SourcePauseTool. During each decision, the tool paused Portal and supplied screenshots, player position and camera angle. Astra selected its inputs before play resumed, so the elapsed test was not equivalent to 23 hours of continuous real-time control.
At list prices, the run’s usage would have cost at least $570, although the developer used a $200 Codex subscription. That gap and the specialized interface matter when comparing the demonstration with ordinary gameplay by a person.
Completing a full 3D puzzle game shows sustained planning across a long sequence of actions, but one successful run is not a general benchmark of computer control. The pausing mechanism, structured state information and game-specific tooling all shaped the task, and performance in less controlled desktop or robotic environments remains a separate question.