OpenAI's GPT-6 Astra Plays Minecraft for 141 Hours, Gets Stuck Farming Potatoes After Creeper Blows Up Its Base

US-based AI evaluation group Vals AI ran a long autonomous gameplay test of OpenAI's newest model GPT-6 Astra in Minecraft, streamed live on Twitch for 141 hours. The test was independently conducted by a third party to observe how the model performs outside a closed testing environment.

GPT-6 Astra showed strong long-term planning, successfully building a semi-automatic Blaze farm, collecting 6 Blaze Rods, hunting Endermen in the Nether's Warped Forest for 3 Ender Pearls, and storing rare items in a camp chest.

But a Creeper sneaked up and exploded, destroying the storage chest and respawn bed. Over a hundred hours of progress was wiped out. After the blast, the model became visibly withdrawn, spending hours just farming potatoes on loop, treating anything green as a threat, even mistaking sugar cane for a Creeper.

The experiment shows the model can handle complex long tasks, but lacks a recovery strategy after sudden losses. Researchers are still debating whether this is emotional simulation or just goal-function degradation after negative feedback, with no clear conclusion yet.