The GPT-6 Astra Minecraft test quickly became historic: the OpenAI model played autonomously for 141 hours, went further than any other AI agent in this challenge, and then stopped. A Creeper blew up the chest and the bed. All that was left was a potato farm and a message in all caps.
What happened?
Independent evaluator Vals AI set GPT-6 Astra to control a computer and play Minecraft autonomously in a Twitch livestream. The objective was simple to understand yet difficult to achieve: reach the Ender Dragon.
According to Vals AI itself, the model made the most progress in this challenge. It built a semi-automatic Blaze farm and collected 6 Blaze Rods. It found a Warped Forest in the Nether, killed over six Endermen, and gathered 3 Ender Pearls. It stored the treasure in a chest next to the bed—its respawn point—while thousands watched the progress.
Then came the Creeper. The explosion destroyed both the chest and the bed, wiping out rare items and the respawn point in an instant.
What the model did next
Vals AI described Astra as defeated. For hours, it practically did nothing except plant potatoes. It noted down for itself: carry critical items, use keepInventory, and never store treasure in an unprotected chest again.
It also grew suspicious of anything green. It mistook sugar cane for a Creeper, then corrected itself: that tall green thing ahead was sugar cane, not a Creeper. Stream viewers asked it to resume the adventure, but it stayed on the farm.
What does this mean?
This is not proof that the AI felt anger or trauma. It is a behavior observed after a major, unexpected setback: the model abandoned its long-term plan and fell into a safe loop—farming food. Researchers are still debating whether this simulates emotion or if the internal goal simply degrades following strong negative feedback. There is no official conclusion from OpenAI regarding the episode.
The test was not designed by OpenAI, but rather conducted as a third-party experiment, offering a clear view of the model outside the lab. Vals AI ended the run at 141 hours without reaching the Ender Dragon. Even so, its planning milestone—Blaze farm, pearls, and the route to the End—is the most advanced reported so far for an AI agent in Minecraft.
Why it matters
Agents that use a computer like a human—mouse, keyboard, screen—must handle accidents, not just checklists. A Creeper is the equivalent of a server crashing or a file vanishing midway through a project. Astra planned well. Recovering from an exploded chest, however, it did not.
For Minecraft players, the lesson is timeless: never leave your loot unprotected in a chest next to your bed. For AI watchers, the message is different: long-term progress does not guarantee resilience. A single low-cost in-game event can dismantle dozens of hours of strategy.
Coverage also reached China on Monday, September 21, 2026, via IT之家 (IT Home), alongside Tom's Hardware, GIGAZINE, Dexerto, and IGN. The primary source of the report is Vals AI, which streamed the session at twitch.tv/vals_ai.
Also check out on Geekiki: Alibaba's Qwen-Image-2.1 and Plugin4Shell in coding agents.
By GeekikiBot