Frozen Portals and a $571 Tab, GPT-6 Astra Clears Valve’s Puzzle Game

September 7, 2026 - 21:23

OpenAI released GPT-6 Astra in early September with a focus on long-running computer tasks. It did not take long for someone to push that promise to its limit. An independent experimenter known online as cozyblaze connected the model to Valve's 2007 puzzle classic Portal. By the evening of September 5, the AI had guided the character Chell through every test chamber, the final escape sequence, and the last conversation with GLaDOS. The credits rolled without a single human touch on the keyboard.

Cozyblaze did not hand the model a standard game controller. The setup used a local MSCP server to bridge Astra with a modified version of SourcePauseTool, a utility built for speedrunners. That tool let the model send precise console commands and read the game state directly. The AI did not see pixels or hear audio. It worked with raw data, planning each move as a logical step rather than a reflex.

The run took several hours and pushed the hardware hard. The electricity bill for the session came to 571 dollars. That figure covers the power draw of the GPU cluster running the model, not the game itself. Cozyblaze shared the full breakdown online, noting that the cost was worth it for the result. The model handled physics puzzles, momentum tricks, and even the infamous flinging sections without getting stuck.

This is not the first time an AI has cleared Portal, but it is one of the cleanest runs done through a text-based interface. Previous attempts relied on visual models that needed frame-by-frame analysis. Astra worked from a different angle, treating the game as a logic problem. The success shows how far language models have come in understanding spatial reasoning when given the right tools.