GPT-5.6 Sol can beat Factorio on peaceful mode

AI agents can now autonomously complete complex, long-term tasks with evolving strategies, like playing Factorio. This signifies that agents are moving beyond simple command execution to sophisticated planning and adaptation.

Ryan Madden reported that GPT-5.6 Sol, operating within the Codex framework with a low thinking level, successfully completed a peaceful mode game of Factorio in 13.5 hours of in-game time. This achievement involved using a game manager and a Python library to enable programmatic interaction with the game, while adhering to the same limitations as a human player using the graphical user interface, such as reach and fog of war. The run demonstrated both surprising strengths and notable weaknesses in Sol's performance. Initially planned to conclude after blue science research, the AI's effectiveness led to extending the game, though scaling issues later emerged, significantly slowing progress. The eventual rocket launch required substantial manual intervention from Sol to compensate for failures in connecting production systems, by ferrying supplies between factory sections. This marks an improvement over GPT-5.5, which struggled with basic mechanics like transport belts.

GPT-5.6 Sol's successful Factorio run indicates a significant advancement in the long-term task completion capabilities of AI agents. The ability to autonomously navigate complex, evolving environments, manage resources, and adapt strategies over extended periods suggests a move towards more sophisticated artificial general intelligence.

This breakthrough in autonomous gameplay offers a promising new benchmark for evaluating AI agent performance, particularly in tasks that mirror real-world programming and complex problem-solving. The Factorio testbed, with its clear objectives and visual feedback, provides a more legible environment for understanding AI decision-making compared to abstract code generation.