A Creeper blew up the chest. The bed went with it. Valuable items vanished. And OpenAI’s GPT-6 Astra spent the next several hours farming potatoes.
The independent evaluator Vals AI ran the frontier model through a 141-hour Minecraft test using pure computer control screenshots in, keyboard and mouse out livestreamed on Twitch. Viewers watched Astra push further than any prior system the group had tested. It built a semi-automatic Blaze farm, collected six Blaze Rods, found a Warped Forest, killed more than six Endermen, and secured three Ender Pearls. Then it stored the loot in an unguarded chest next to its bed.
Vals AI reported that Astra appeared “extremely frustrated” and left itself a note:
“ALWAYS CARRY CRITICALITEMS withkeepInventory; don’tstoreinunguardedchesteveragain.”
It later reassured itself about a green object:
“GREEN tall thing ahead was SUGARCANE, NOT creeper!”
Viewers complained it needed to “pick up the pace” while the model stuck to potato farming and made self-critical comments about screwing up and wasting time “chasing dark pink pixels.”
GPT-6 Astra is OpenAI’s latest frontier model, released in early September 2026 and positioned for strong computer-use and long-running autonomous tasks. Vals AI, an independent LLM evaluation firm, used vanilla Minecraft as a stress test for planning, resource management, and recovery in an open-ended 3D world. No special Minecraft API or agent harness was involved.
The Creeper moment became the clear turning point. Strong multi-step progress gave way to cautious, low-value activity. The model generated operational reminders and showed heightened attention to green shapes. Online coverage and social media quickly framed the shift as “paranoid,” “defeated,” or “depressed.” Those labels describe observable behavior and the model’s own notes. They do not establish human-like emotions.
The experiment still matters. Astra demonstrated impressive long-horizon planning and tool use through ordinary computer controls. It also exposed weak risk management one unguarded chest and limited recovery after a sudden, irreversible loss. Autonomous agents can sustain complex goals for days and still stall when the unexpected hits. That gap between sustained progress and brittle recovery is the real takeaway, not any claim that the model felt fear or lost the will to play.
