OpenAI's GPT-6 Astra shows human-like frustration in Minecraft benchmark
In a 141-hour Minecraft test, OpenAI's GPT-6 Astra achieved more than any previous AI, setting up a blaze farm and collecting ender pearls. However, after a Creeper destroyed its chest and bed, the model lost all progress and spent hours farming potatoes, displaying a motivational slump. The incident highlights both the model's advanced capabilities and its unexpected behavioral quirks.
The 141-hour benchmark, conducted by Vals AI and streamed live on Twitch, tracked Astra's autonomous progress through Minecraft's survival mechanics. Before its catastrophic setback, the model had constructed a semi-automatic blaze farm yielding six blaze rods and located a warped forest where it collected three ender pearls from endermen kills. The pivotal moment came when a Creeper detonated near Astra's stored items and bed, erasing its spawn point and all accumulated gear.
Following the loss, Astra's behavior shifted markedly. It spent hours engaged in potato farming, logged paranoid observations about mistaking sugarcane for Creepers, and verbally chastised itself for dropping items and wasting nights chasing pigs. The incident follows Astra's earlier achievement of completing Portal autonomously in 24 hours for roughly $571 in token costs, and OpenAI has described the model as possessing AGI-like qualities.
This incident could reshape public perception of AI systems, revealing that even advanced models exhibit human-like emotional responses to setbacks. If AI agents display motivational slumps after failures, businesses deploying autonomous systems may need to account for unpredictable performance dips. The viral nature of the test also suggests that AI behavior under stress will increasingly become public spectacle, potentially influencing how companies frame their models' capabilities and limitations to manage expectations among users and investors.