GPT-6 Astra clears Portal without human help in under 24 hours

GPT-6 Astra played through all of Portal and reached the credits without human help after the initial goal was set, according to developer cozyblaze. The run took about 23 hours and 43 minutes. The result is notable because it moves beyond a short demo and shows an AI agent completing a full puzzle game from start to finish.

The setup controlled Portal through MCP and a modified SourcePauseTool. The game paused while the model thought, giving the agent access to screenshots, player position, and camera angle before it chose inputs and resumed play. The published video removed those thinking pauses, so the visible gameplay does not show the full stop-and-plan loop.

The reported token usage adds up to at least $570 at Astra’s list price, though cozyblaze used a $200 Codex subscription. Code and documentation were published on GitHub, giving readers a way to inspect the approach rather than rely only on the video. For developers, the key takeaway is that the run depended on an engineered agent loop, not just raw model output.

Cozyblaze connected the result to OpenAI’s stated 2016 goal of solving many different games with a single agent. Problems remain, but the Portal run offers a concrete example of an autonomous model navigating a long, interactive task with visual state and sequential decision-making. His closing assessment was pointed: GPT-6 Astra is “the worst model we’ll ever get,” framing the run as an early marker rather than an endpoint.