动察 Beating AI Flash: Independent developer cozyblaze has hooked GPT-6 Astra into Valve's 3D puzzle game Portal. At the start, he gave it just one instruction: play all the way to the end credits on its own, without going online to look up walkthroughs. Astra ultimately succeeded in beating the game, taking approximately 23 hours and 43 minutes in total, consuming around 435 million tokens, which translates to roughly $571 at API pricing.

Astra plays games differently from humans. When it thinks, the game pauses, and it then examines the current screen, character position, and viewpoint to decide where to move next, where to look, and where to place portals, before letting the game resume. Along the way, OpenAI's service experienced capacity issues, and the developer manually restored the task, later switching to the faster Fast mode. However, public logs show that from the start to the appearance of the end credits, he provided no further gameplay hints.

That said, this is not a rigorous benchmark. Portal has been out for nearly 20 years, with a wealth of guides and videos online, and even though the experiment barred Astra from actively searching for walkthroughs, it cannot rule out that relevant game knowledge has already entered its training data. What's truly intriguing is that AI can now carry out a task lasting nearly a full day, requiring continuous observation and manipulation of a 3D environment, all the way to completion.