GPT-5.6 Sol Beats Factorio in 13.5 Hours, Launches Rocket

OpenAI's GPT-5.6 Sol has achieved a notable milestone in autonomous gaming, completing a peaceful Factorio run in 13.5 game hours. Running in Codex at low thinking level, the model navigated the complex factory-building simulation from initial resource gathering through rocket launch, demonstrating improved strategic planning over its GPT-5.5 predecessor. Researcher Ryan Madden documented the full run with production statistics and video recordings, establishing a reusable framework for evaluating model capabilities through gameplay.

Gameplay Setup and Constraints

The experiment used a Python-based game manager that allowed Sol to interact with Factorio programmatically while enforcing realistic human player constraints. The system implemented reach limitations and fog-of-war mechanics, preventing the model from accessing information or performing actions outside normal gameplay boundaries. This setup ensures the capability test reflects genuine decision-making and planning rather than exploiting programmatic advantages unavailable to human players.

According to Madden, the peaceful mode configuration removed enemy biters from the game, isolating the challenge to resource management, logistics design, and production optimization. The model controlled all aspects of factory construction, from mining operations through automated science production systems required for rocket launch.

Performance Milestones and Improvements

Sol demonstrated clear advancement over previous GPT-5.5 attempts, which failed to utilize transport belts effectively. The October 2026 run showed competent belt usage throughout early and mid-game phases, successfully automating red and green science production before scaling to blue science research. Production statistics captured throughout the run reveal steady resource accumulation aligned with research milestone achievements.

The model exceeded initial expectations during the blue science phase, prompting Madden to continue the run toward rocket launch rather than terminating at the originally planned stopping point. Video documentation shows Sol constructing organized production chains and managing multiple simultaneous manufacturing processes across different factory sectors.

Late-Game Scaling Challenges

Despite strong early performance, Sol encountered significant scaling difficulties in late-game production. The model struggled to establish reliable automated connections between distant factory sections, instead resorting to repeated manual transportation of supplies. This pattern of ferrying materials by hand rather than building robust logistics infrastructure slowed progress substantially during the final phases before rocket launch.

These scaling pains appeared later than Madden anticipated, emerging primarily after blue science completion rather than during earlier production stages. The behavior suggests current limitations in long-term planning and system integration when managing increasingly complex factory layouts.

Release Date and Availability

This capability demonstration was officially released on October 31, 2026, with complete run documentation including YouTube timelapse footage and annotated production charts. The evaluation framework and recorded gameplay data provide a publicly accessible benchmark for comparing future model versions and alternative AI systems.

Evaluation Framework Implications

Madden emphasizes Factorio's potential as a medium for building intuition about model capabilities, noting the game's programming-analogous nature while remaining visually legible to casual observers. This accessibility makes gameplay recordings effective for grounding discussions of model behaviors without requiring technical expertise to interpret results.

• Full run recorded with screencast and production statistics
• Python game manager code establishes reusable evaluation harness
• Future tests planned for other models, thinking levels, and non-peaceful modes
• Video commentary in development to analyze specific decision patterns

The researcher plans to test additional models and configurations, including combat scenarios with enemy biters enabled. These expanded evaluations will help map the boundaries of current AI capabilities in complex, dynamic environments requiring sustained strategic thinking.

What This Means

GPT-5.6 Sol's successful Factorio completion marks measurable progress in autonomous agent capabilities, particularly in sustained task execution requiring multi-step planning and resource optimization. While late-game scaling issues reveal current limitations, the 13.5-hour run represents a significant advancement over previous model generations. The established evaluation framework offers researchers a standardized, interpretable method for assessing AI system performance in domains requiring both tactical execution and strategic foresight, with clear applications to software development, logistics optimization, and complex project management scenarios.