OpenAI's GPT-6 Astra beats Portal solo in under 24 hours

The feat is a concrete step toward OpenAI's long-stated ambition of a single agent mastering many games, a vision the developer cozyblaze explicitly referenced. By finishing a full, unmodified puzzle game like Portal, the model demonstrates planning and decision-making that goes beyond narrow single-task benchmarks, even though it relied on a tool to pause the game while it reasoned.

At the same time, the cost and complexity underscore how far general game-playing AI is from being practical. The run burned at least $570 worth of tokens at list price, and required a custom pause tool to give the model time to think. That suggests that while autonomy is advancing, efficiency and the ability to act in real time remain hurdles. The developer's comment that GPT-6 Astra is "the worst model we'll ever get" is a reminder that this capability will likely look primitive compared with what comes next, and that progress is expected to continue.

The significance hinges on whether such autonomous play can be generalized to other, less structured tasks. If this approach scales and becomes cheaper, it could point toward AI agents that handle long-horizon goals in software environments with minimal human input. For now, the achievement is a proof of concept, with the 23-hour runtime and token cost serving as markers of how far there is to go.