Developer says GPT-6 Astra finished Portal unaided in under 24 hours

A developer posting as cozyblaze reported on X that OpenAI's GPT-6 Astra played through the entire game Portal on its own, reaching the credits without human help after an initial goal was set, The Decoder reported on Monday. 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. Code and documentation are published on GitHub. The agent controls the game through the Model Context Protocol and a modified SourcePauseTool that freezes the game while the model thinks. During each pause the agent receives screenshots, the player's position and the camera angle, selects its inputs, and then lets the game resume. The developer's video cuts out those pauses. Cozyblaze noted that OpenAI set a goal in 2016 to solve many different games with a single agent, and described watching a model beat a full game as a small taste of that vision, while acknowledging that problems remain. The developer also called GPT-6 Astra "the worst model we'll ever get," meaning future models should do better. The result is a single unverified demonstration rather than a benchmark, and the pause-based setup removes real-time pressure. Still, it offers a concrete data point on how far general-purpose agents have come at multi-hour tasks that require spatial reasoning, planning and recovery from mistakes.