GPT-5.6 Sol can beat Factorio on peaceful mode

What happened

A developer had GPT-5.6 Sol, running in Codex, play Factorio in peaceful mode. The AI played through a small code library that let it control the game with Python. The library followed the same rules a human player faces in the normal game interface, such as limited reach and fog of war (areas the player cannot see yet). The full run is shown in a timelapse video on YouTube.

Key numbers and facts

• The run took 13.5 hours of game time.
• The author first planned to stop after blue science research, but kept going.
• The rocket launch happened, but the AI used a lot of manual work to get there.
• The author compared this run to earlier attempts with GPT-5.5, which could not use transport belts (conveyor belts that move items) at all.

Background

Factorio is a game where players build factories that make items and science. Earlier posts by the same author tested Claude Code and other AI agents on the game. The author uses Factorio as a test because it is easy for people to watch and understand how the AI is doing.

Why it matters

The run shows that AI can plan and build a large, working factory over many hours. It also shows clear limits. Late in the game, the AI failed to connect some production systems and hid this by carrying supplies by hand between factory areas. The author thinks Factorio is a useful way to understand what AI models can and cannot do, since it is similar to programming but easier to watch.

What to watch next

• A more detailed commentary video and highlights post from the author.
• Tests of other AI models and thinking levels.
• Runs with biters turned on, which are harder than peaceful runs.