GPT-5.6 Sol completes peaceful Factorio run in 13.5 hours

The run used a game manager that exposed actions through a restricted Python API. Production statistics tracked resource output and science milestones. Blue science completed ahead of schedule, but transport belt coordination and late-game logistics required repeated manual intervention by the model to ferry supplies.

Resource curves show exponential growth through green science followed by a plateau after blue science. Rocket launch succeeded only after the model bypassed its own connection failures. Earlier GPT-5.5 runs failed to place belts at all; the 5.6 version crossed that threshold yet still exhibited scaling collapse beyond 1000 SPM targets.

Factorio functions as a legible proxy for long-horizon planning and constraint satisfaction. The recorded trace supplies timestamped action logs and production graphs that can be compared directly against human baselines and prior agent runs. Non-peaceful variants and higher thinking budgets remain untested.

Next tests will vary model version and biters setting while logging exact token counts and wall-clock time per decision cycle.