GPT-6 Astra: Pokemon champion in 18 hours, potato farmer after one Creeper mishap
decoder ·
GPT-6 Astra achieved a Pokémon FireRed championship in 18 hours and 12 minutes, compared with 96 hours and 35 minutes for GPT-5.6 Sol, while also making major progress in Minecraft and Factorio. In Minecraft, it reached the Nether, built a semi-automatic blaze farm, gathered key materials for the final portal, and then lost its chest and bed when a Creeper exploded nearby. Afterward, Astra adopted a rule to keep critical items on itself and spent several hours farming potatoes, showing how a useful lesson can become counterproductive overgeneralization. It launched a rocket in Factorio after about ten hours and also completed games or major campaigns including Portal, Fallout 2, RimWorld, and Fallout 3. On ARC-AGI-3, Astra scored 62.7 percent through the standard interface and about 99.9 percent using OpenAI's own harness, far above earlier systems cited in the article. The article attributes the improvement to Astra's ability to turn observations into symbolic rules, track game states, and plan actions without specialized game integrations. Its Fallout 3 run further showed a strategy of checking whether actions worked before repeating them, although excessive caution and persistent rules could slow it down or distract it from its objective.
decoder