GPT-5.6 Sol Launched a Factorio Rocket on Peaceful Mode, With Hauling Done by Hand

OpenAI's GPT-5.6 Sol launched a rocket in Factorio on peaceful mode, Ryan Madden reported on October 31. He wrote that the launch took "significant 'manual' effort" from the model.

Madden, who runs the blog ryanmadden.net, ran Sol in Codex at the "low" thinking level. The run covered 13.5 hours of in-game time, according to his post. It does not report wall-clock time, token use or cost.

How the run was set up

A game manager let the model act through a small Python library. Madden wrote that the library enforces the limits a human faces in the graphical client, such as player reach and fog of war.

Peaceful mode removes the biter enemies. Madden's post covers only that setting, and he wrote that he plans to test other models, other thinking levels and non-peaceful runs.

The post describes one run. It does not mention repeats, so there is no success rate.

What the model did by hand

Madden wrote that Sol repeatedly carried supplies between factory areas because it had not connected some production systems. Late-game scaling problems appeared later than he expected and slowed progress.

He had planned to stop after blue science research. He kept going because Sol was doing better than he expected.

He wrote that GPT-5.5 failed to use transport belts at all. His post gives no score for either model.

What came before

Madden's April 5 post on Claude Code used a command-line tool with about 20 commands. Claude, whose model version that post does not name, made it through red and green science and was halfway through "automation 2" research when the second run ended.

That setup let the agent reach any entity without walking and mine without delay. Madden wrote that this likely pushed it toward a crate-assembler-crate pattern and away from belts. The October library enforces reach limits, so the two runs are not a like-for-like comparison.

Run	Date	Harness	Result
Claude Code	Apr. 5	Bash CLI, about 20 commands	Stopped mid automation 2
GPT-5.5	Oct. 31 post	Python library	Did not use belts
GPT-5.6 Sol, low	Oct. 31	Python library	Rocket launched, peaceful

How hard the benchmark was meant to be

The Factorio Learning Environment paper, submitted March 6, 2025 by Jack Hopkins, Mart Bakler and Akbir Khan, said one rocket component needs more than 60 interlinked machines making over 1,000 items a minute. The best lab-play result it reported was Claude 3.5 Sonnet at 21.9% task success.

The paper's authors listed trouble coordinating more than about six machines on items with over three ingredients. They also warned that their Python interface differs from playing with a mouse, and that humans were not tested through the same API.

Madden wrote that he is "optimistic about autonomous Factorio play as a medium for building intuition about the nature of model capabilities." The paper's lab-play trajectories ran 128 API calls each. Madden's post reports no step count for the rocket run.

For other recent agent work on Claude Code, see Anthropic's rule against recreating existing games and a heuristic compiler reanimated with Claude Code.