Ryan Madden guided GPT-5.6 Sol through a peaceful Factorio session in Codex at a low thinking level, finishing a 13.5-hour game-time run that reached rocket launch. A Python game manager enforced human-style constraints, and production charts show scaling failures hitting after blue science.
Ryan Madden guided GPT-5.6 Sol through a peaceful Factorio session inside Codex at a low thinking level. The model finished in 13.5 hours of game time, launching a rocket after Madden extended the run past the blue science stage.
A Python game manager handled programmatic interaction, enforcing the same limits human players face: restricted reach and fog of war. Madden captured production statistics throughout and annotated science milestones on a total-resource chart.
The chart tracks resource output climbing through early and mid-game milestones. Madden intended to stop after blue science, but Sol’s early factory layout convinced him to continue. Scaling issues arrived later than expected. Sol failed to connect late-game production systems, so it ferried supplies between zones to keep rocket assembly running.
Sol handled early logistics better than previous attempts. GPT-5.5 never used transport belts; this version built belts and organized factories, though it collapsed when systems needed to connect at scale. Madden calls autonomous Factorio a legible test bed for model reasoning. The game mirrors programming but produces visible results anyone can read. He plans a commentary video over the full recording and will test other models, thinking levels, and non-peaceful runs.