GPT-5.6 Sol completes a peaceful Factorio run in 13.5 hours

Beating the factory game

The GPT-5.6 Sol model recently achieved a milestone by completing a peaceful run of Factorio. Factorio is a complex automation game where players must build factories to produce resources and conduct scientific research. This specific run lasted 13.5 hours of game time, marking a significant step forward from previous attempts.

In earlier tests, models like GPT-5.5 struggled with basic mechanics, such as using transport belts to move items. The new run with Sol demonstrated that the model can now handle the fundamental logistics required to build a functional industrial base. This success allows researchers to see how the AI handles long-term goals and spatial reasoning.

The run was conducted on a peaceful setting, which means the AI did not have to defend against hostile creatures called biters. By removing combat, the test focused entirely on the model's ability to plan, build, and manage a growing production network. This isolation helps clarify the model's specific strengths in logic and construction.

How the AI interacts

To enable the AI to play, a game manager was used to bridge the gap between the model and the software. The model, running in Codex with a low thinking level, interacted with the game through a small Python library. This setup allowed the AI to execute actions programmatically rather than clicking a mouse like a human would.

The system was designed to enforce the same restrictions that a human player faces when using the game's graphical user interface. For example, the AI was limited by the character's reach and the fog of war, which prevents the player from seeing distant areas. These constraints ensure that the AI's performance is a true reflection of its problem-solving capabilities.

By using a programmatic interface, the researchers can capture precise data on the AI's decisions. This includes production statistics and research milestones that are easier to analyze than a standard video feed. This method provides a clear window into how the model prioritizes different tasks over many hours of operation.

Successes and bottlenecks

The run showed that Sol was surprisingly effective during the early and middle stages of the game. It successfully navigated the requirements for blue science research, which initially exceeded the expectations of the researchers. The model managed to set up complex production chains that kept the factory moving forward steadily.

However, the AI eventually hit a wall when the game's demands began to scale up significantly. While the early stages were smooth, the late-game requirements caused the model's relative progress to slow down. These scaling issues made it difficult for the AI to maintain the efficiency it displayed during the beginning of the run.

To reach the final goal of a rocket launch, the AI had to perform what the researcher described as 'manual' effort. Because it struggled to connect certain production systems perfectly, it repeatedly ferried supplies between different areas of the factory. This behavior acted as a workaround for the model's inability to solve the complex logistics of the late-game environment.

A new testbed for AI

The researchers view Factorio as a valuable medium for building intuition about the nature of model capabilities. Because the game is analogous to programming, it provides a structured way to observe how an AI thinks through a series of instructions. It offers a level of legibility that makes it easier for observers to understand the model's internal logic.

This experiment serves as a way to ground discussions about AI behavior in a visible, interactive environment. Instead of just looking at code output, people can watch a factory grow and see where the AI's planning succeeds or fails. It provides a tangible way to measure the limits of autonomous agents in complex environments.

Moving forward, the goal is to test other models and different thinking levels to see how they compare. The researchers also plan to test non-peaceful runs to see how the AI handles dynamic threats. These tests will help determine how well AI can handle unpredictable variables alongside complex construction tasks.