Experiment: GPT-6 Astra Plays Portal Autonomously
Independent developer cozyblaze hooked OpenAI's GPT-6 Astra into Valve's 3D puzzle game Portal. He gave the model a single instruction: play through the game all the way to the end credits without searching for any walkthroughs online.
Time and Cost
Astra eventually completed the game, taking approximately 23 hours and 43 minutes. During the playthrough, it consumed about 435 million tokens, which at OpenAI's API pricing translates to roughly $571.
How Astra Played
Astra's approach differed from human gameplay. The game would pause while Astra thought, analyzing the current screen, the player's position, and the camera angle. It would then decide the next action — how to move, where to look, and where to place portals — before resuming the game. Along the way, OpenAI's service experienced capacity issues, and the developer manually resumed the task. He later switched to the faster "Fast" mode. According to public logs, the developer did not provide any game hints from the start to the appearance of the end credits.
Limitations
This is not a rigorous benchmark. Portal has been available for nearly 20 years, and abundant walkthroughs and videos exist online. Although the experiment prohibited Astra from actively searching for guides, it's possible that related game knowledge was already included in its training data. The noteworthy aspect is that AI can now sustain a task lasting nearly a full day, involving continuous observation and manipulation of a 3D environment.