Ryan Madden has demonstrated that the GPT-5.6 Sol model, running in a Codex environment with a low thinking level, can successfully complete a peaceful run of Factorio—a complex automation and factory-building game. The model achieved this milestone in 13.5 hours of in-game time. To facilitate this, Madden utilized a game manager that allowed the model to interact with the game programmatically via a Python library. This setup was designed to ensure the AI faced the same physical limitations as human players, including restricted movement reach and the 'fog of war' mechanic, which hides parts of the map not yet explored.

A rocket launching into the atmosphere in the game Factorio.
The GPT-5.6 Sol model successfully reached the final stage of the Factorio run, culminating in a rocket launch.ryanmadden.net

The image depicts the successful completion of the Factorio run, which is the primary milestone achieved by the GPT-5.6 Sol model.

The performance of the model showed a mix of high-level competence and specific mechanical failures. While Sol demonstrated a surprising ability to manage production and reach science research milestones, it struggled with the core logistics of the game. Specifically, the model frequently failed to utilize transport belts to route resources, opting instead to store items in chests and manually ferry supplies between different areas of the factory. This 'manual' effort was particularly evident during the final stages of the run, where the model had to repeatedly move supplies to connect production systems and eventually achieve a rocket launch.

A dashboard showing production statistics and resource flow in Factorio.
Production statistics from the Factorio run demonstrate the model's ability to manage complex factory systems.ryanmadden.net

This image illustrates the high-level competence of the model in managing production and reaching science research milestones.

Despite these logistical hurdles, the run represented a significant improvement over previous attempts with GPT-5.5, which had failed to use transport belts entirely. Madden noted that while the model eventually hit scaling issues during the late-game science stages, its ability to navigate the factory's complexities provided valuable insight into current model capabilities. The experiment highlights Factorio as a legible medium for testing AI intuition, as the visual nature of the factory makes it easier for observers to identify exactly where a model's logic succeeds or fails compared to traditional programming tasks.