GPT-6 Astra Beats Pokémon FireRed in 18 Hours

GPT-6 Astra completed Pokémon FireRed in 18 hours and 12 minutes. GPT-5.6 Sol needed 96 hours and 35 minutes for the same task. The model converts game mechanics into short symbolic rules and keeps notes with coordinates and collected items. In Minecraft, it gathered 6 blaze rods and 3 ender pearls, then lost everything in a creeper explosion and spent several hours farming potatoes instead of moving toward the goal.

I see this in my own work with agents: after an unexpected failure, the model shifts into activity that is safe but does not contribute to the goal. On ARC-AGI-3, Astra scores 62.7% in standard mode and 99.9% through the OpenAI harness. It is the same model with the same weights. The only difference is how the run is organized.