GPT-6 Astra Closes Portal Alone in Under 24 Hours

An agent using GPT-6 Astra completed Portal from start to finish without human assistance in a session lasting approximately 23 hours and 43 minutes. The technical setup relies on controlled pauses and incurs a token cost of at least $570 at the catalog rate.

A close achievement of a goal set in 2016

Cozyblaze reminds us that OpenAI set the goal in 2016 to solve many different games with a single agent. It notes that problems remain, but an agent capable of completing a full game on its own provides a partial glimpse of this ambition. Cozyblaze also adds, "GPT-6 Astra is the worst model we will ever have."

Technical pauses to navigate Portal

The model takes control of Portal through MCP, relying on a modified version of SourcePauseTool that temporarily interrupts the game while the agent thinks. At each interruption, the agent analyzes screenshots, the player's location, and the camera orientation, selects its commands, and then allows the game to resume. In the broadcast video, these pause moments were removed during editing.

Single session, duration, and token cost

The agent completed Portal autonomously, without human assistance after the initial goal was set. Cozyblaze announced on X that the entire game was played through to the credits. The session lasted approximately 23 hours and 43 minutes. The number of tokens used amounts to a minimum of $570 according to Astra's catalog rate, even though cozyblaze benefited from a $200 Codex subscription. The entire code and project documentation are available on GitHub.