The GPT-6 Astra model has demonstrated significant progress in autonomous gaming, completing tasks that previously stumped AI systems. In Pokemon FireRed, Astra achieved the champion title in 18 hours and 12 minutes, whereas GPT-5.6 Sol required 96 hours and 35 minutes.
In Minecraft, Astra reached further than previous models, building a semi-automatic blaze farm and gathering ingredients for eyes of ender. The run lasted 141 hours before being called off by Vals AI. The agent operates via screen, mouse, and keyboard inputs rather than specialized game hooks.
Astra also completed Fallout 3 in approximately 59 hours, navigating the game's story and mechanics autonomously. In Factorio: Space Age, the model produced blue science packs in two hours and launched its first rocket in roughly ten hours.
The model's performance on the ARC-AGI-3 benchmark reached 62.7 percent using a standard interface and 99.9 percent with OpenAI's harness. This compares to 7.78 percent for GPT-5.6 Sol and over 30 percent for Claude Opus 5.
Astra utilizes a method of translating game mechanics into symbolic descriptions to track objects and rules. However, the model can overcorrect; after a Creeper destroyed its items in Minecraft, it established a rigid rule to avoid unguarded chests, subsequently spending hours farming potatoes.
