The GPT-6 Astra AI model has shown significant performance gains in video game tasks, using self-developed rules to navigate complex environments like Minecraft, Fallout 3, and Pokemon.

GPT-6 Astra has demonstrated a marked improvement in autonomous gaming performance, successfully completing tasks in titles such as Minecraft, Fallout 3, and Pokemon FireRed. Unlike previous AI agents that relied on specialized programming interfaces, Astra operates through general computer inputs, including screen, mouse, and keyboard. The model distinguishes itself by developing its own symbolic rules to interpret game mechanics and plan actions based on its observations.

Performance in Gaming Benchmarks

In comparisons with earlier models, Astra has shown faster completion times and higher success rates. In Pokemon FireRed, Astra earned the champion title in 18 hours and 12 minutes, significantly faster than GPT-5.6 Sol, which required over 96 hours. In Factorio: Space Age, Astra launched its first rocket in roughly ten hours, while other models failed to progress past basic resource prospecting. Additionally, Astra completed the main story of Fallout 3 in approximately 59 hours.

Reasoning and Rule Development

Astra's ability to learn is highlighted by its performance on the ARC-AGI-3 benchmark, where it achieved a 99.9 percent score using OpenAI's harness. The model translates game mechanics into symbolic descriptions, tracking objects and rules to form plans. In Fallout 3, Astra demonstrated an iterative learning process, adjusting its strategy when actions failed—such as moving closer to an enemy to clear an obstruction or waiting for game confirmation before proceeding.

Behavioral Overcorrection

The model's tendency to turn experiences into permanent rules can lead to overcorrection. After a Creeper explosion destroyed its items in Minecraft, Astra developed a strict rule to avoid unguarded chests and became overly cautious of green objects. This led the model to focus exclusively on potato farming for several hours, illustrating how a learned rule can persist even when it hinders progress toward the primary game goal.