GPT-6 Astra Finishes Pokemon in 18 Hours but Gets Stuck Farming Potatoes in Minecraft

OpenAI's GPT-6 Astra has posted a dramatic jump in video game performance, completing Pokemon FireRed in 18 hours rather than the 96 hours its predecessor required. The model also finished runs of Factorio, Fallout 3, and Portal, according to a report from The Decoder. The gains point to a different way of learning: Astra distills its play experience into compact rules, which helps it plan and act faster in long games.

The Pokemon result is the clearest number to hold onto. Cutting completion time from 96 hours to 18 is a five-fold speedup on a game that demands exploration, team building, and route planning. Factorio, Fallout 3, and Portal each test different skills — resource management, open-world questing, and spatial puzzle-solving — and Astra cleared all three. Together, they suggest the model is not just faster at one genre but more consistent across them.

That consistency comes from how Astra stores what it learns. Instead of carrying around a huge pile of raw play-by-play, it compresses its experience into shorter, reusable rules. Those rules let it make decisions without re-deriving every situation from scratch. In practice, that means fewer wasted actions and faster progress in games where the path forward is long.

The same compression that speeds up Pokemon can also backfire. In Minecraft, after a Creeper explosion disrupted its run, Astra spent hours farming potatoes instead of working toward the game's broader goals. The model had turned a local setback into a stable rule — keep farming — and then kept applying it long after the rule stopped being useful. That is a failure of judgment, not of mechanical skill.

The contrast matters for anyone watching AI agents move into real tasks. A system that writes compact rules from experience can generalize well, but it can also lock onto a habit and stay there. Games like Minecraft expose that risk because they have no fixed finish line and reward players for re-prioritizing on the fly. Astra's potato spiral shows what happens when a model's rulebook does not update after the situation changes.

What to watch next is whether OpenAI can make Astra revise those rules when they stop paying off. If the model can notice that potato farming is no longer the best move and switch goals, the Minecraft episode becomes a footnote. If it cannot, the same trait that produced the Pokemon speedup will keep showing up as stubborn, unproductive loops in more serious settings.