GPT-6 Astra Completed Portal on its Own in 24 Hours: But It Came at a High Cost

OpenAI’s new flagship model, GPT-6 Astra, completed Valve’s 3D puzzle game Portal from start to finish without human intervention in a remarkable experiment. The experiment clearly demonstrated that large language models can go beyond just generating text, evaluating visual information to make decisions and take action in the game world.

It Cost $570

According to the results shared by AI enthusiast CozyBlaze, who conducted the experiment, Astra made a total of 3,336 tool calls while playing Portal. The cost of the experiment, calculated via API, was $571.18. However, CozyBlaze stated that this cost was not directly out of pocket but was covered under a $200 Codex Pro subscription.

The system's operation differs slightly from Astra continuously controlling the game in real-time. CozyBlaze notes that Portal was controlled via MCP and a modified SourcePauseTool (SPT).

In this setup, the game is kept paused while the model thinks. During this time, Astra is provided with screenshots along with information about the player character's position. After making the necessary evaluation, the model sends a series of input commands. Subsequently, SPT executes the movements determined by Astra.

Completed in 24 Hours

Astra took 24 hours to complete the game. When the thinking processes are removed, it is observed that the game was completed in 2 hours.

The model's process of examining screenshots at each stage, evaluating the character's position, and generating subsequent inputs significantly increases the operating time.

CozyBlaze states that the experiment should not be considered an AI benchmark and that the system still has issues that need to be resolved. Nevertheless, a general-purpose agent independently completing the game from start to finish demonstrates how far OpenAI's 2016 vision of solving different games with a single agent has progressed today.

Meanwhile, the source code for the Portal Agent used in the experiment has been shared on GitHub, along with the necessary instructions for those who wish to examine or try the system.