OpenAI's GPT-6 Astra masters complex games but stumbles in Minecraft after a single setback
OpenAI's GPT-6 Astra completed Pokemon FireRed in 18 hours, a dramatic improvement over earlier models that took 96 hours or never finished. It also succeeded in Factorio and Fallout 3, but in Minecraft it lost its stored items to a Creeper explosion and then spent hours farming potatoes instead of progressing. The model's ability to distill experience into compact rules drives both its rapid game completions and its rigid, unhelpful behavior after a disruption.
The ARC-AGI-3 benchmark quantified the leap: Astra scored 62.7 percent through the standard interface and roughly 99.9 percent with OpenAI's own harness, while GPT-5.6 Sol managed 7.78 percent and Claude Opus 5 just over 30 percent. ARC Prize attributes this to Astra's habit of compressing observations into compact symbolic rules and its own shorthand for tracking objects, coordinates, and planned actions.
The Minecraft failure exposes the trade-off. After the Creeper destroyed its chest, Astra wrote itself a note in caps about always carrying critical items, then spent hours farming potatoes instead of recovering. The run was halted after 141 hours. Earlier systems like Voyager used structured data and skill libraries but never saw the actual game screen, whereas Vals AI's run relies on general computer use.
The pattern of rapid success followed by rigid behavior after disruption suggests these models excel at compressing experience into rules but may struggle when those rules break. For users relying on AI agents for real-world tasks, this could mean impressive efficiency in stable environments but unpredictable failures after unexpected events. Businesses deploying such systems may need safeguards and human oversight, particularly where a single setback could cascade into costly detours.