In a Vals AI experiment testing long-horizon game-playing ability, OpenAI's GPT-6 Astra played Minecraft for 141 hours and 33 minutes. Over the run it built a base, farmed crops, tamed wolves, and reached the Nether while carrying goals across a very long session.

But after a Creeper explosion destroyed the chest holding its collected items and its bed, the model stopped making much progress and spent several hours doing almost nothing but farming potatoes. It also became more cautious about Creepers: when it spotted a tall green object, it explicitly checked that it was sugar cane rather than a Creeper.