GPT-6 Astra has played through all of Portal on its own. Developer cozyblaze posted on X that the model navigated the full game and reached the credits without any human help after the initial goal was set. The run took about 23 hours and 43 minutes. Token usage adds up to at least $570 at Astra's list price, though cozyblaze used a $200 Codex subscription. The model controls Portal through MCP and a modified SourcePauseTool that pauses the game while it thinks. During each pause, the agent sees screenshots, player position, and camera angle, picks its inputs, then lets the game resume. The video below has those pauses cut out.

Code and docs are on GitHub. Cozyblaze noted that OpenAI set a goal in 2016 to solve many different games with a single agent. Problems remain, but watching an agent beat a full game on its own is a small taste of that original vision. GPT-6 Astra is "the worst model we'll ever get," cozyblaze added.

Github