GPT-6 Astra completed the full game Portal in 23 hours and 43 minutes without human intervention, according to developer cozyblaze. The model navigated the game from start to finish using a modified SourcePauseTool and MCP to control the game.
The run cost at least $570 at Astra's list price, though cozyblaze used a $200 Codex subscription to reduce costs. During each pause, the agent saw screenshots, player position, and camera angle, then selected inputs before resuming the game.
The model uses MCP and a modified SourcePauseTool to interact with Portal, pausing the game while it processes inputs. The video of the run has these pauses removed, showing the model's progress through the game.
"GPT-6 Astra is 'the worst model we'll ever get,'" said cozyblaze, developer of the model. The developer noted that while the run is a small step toward OpenAI's 2016 goal of solving many games with a single agent, challenges remain.
The achievement follows OpenAI's long-term vision for AI to solve complex tasks autonomously. Cozyblaze highlighted that while the model completed the game, the broader goals of AI autonomy remain unfulfilled.
Cozyblaze did not comment on future updates or improvements, and the open question remains about the model's ability to handle more complex tasks. The developer noted that the run represents a small taste of the original vision.
Source: thedecoder
