GPT-5.6 Sol launched a Factorio rocket in 13.5 hours, mostly by hand

GPT-5.6 Sol, running in Codex at the low thinking level, launched a rocket in Factorio. Ryan Madden's write-up, posted on 31 October and on the front page of Hacker News that day, puts the peaceful run at 13.5 hours of game time: a milestone, and a narrower one than "beats Factorio" suggests.

What the run actually was

Madden has been pointing coding agents at the game for a while; earlier attempts are titled "We got Claude Code playing Factorio", "Towards beating Factorio autonomously" and "Faketorio: an agent testbed". This time Sol reached the game through a manager plus a small Python library; the previous run used Bash. The library enforces what a human at the GUI faces: player reach, fog of war. Biters were switched off, so nothing attacked the factory while it grew.

He had planned to stop after blue science research, and kept going because Sol was more effective than expected. The scaling problems he expected at that stage appeared later instead, slowing relative progress significantly. He published production charts annotated with research milestones and a 60x timelapse of the whole run.

The caveat is the interesting part

Madden does not claim the factory worked. "The rocket launch was only achieved through significant 'manual' effort on Sol's part as it papered over its late-game failures to connect certain production systems by repeatedly ferrying supplies between areas of the factory."

Read that twice. Ferrying supplies by hand is what a player does when the bus they built never reaches the assembler: it is a workaround for a design the agent could not finish, not proof that it finished one. The win condition was met; the system the win condition stands in for was not. Against GPT-5.5, which he says failed to use transport belts at all, that is progress of a specific and legible kind.

One map, one seed, no biters

This is a capability demo, not a benchmark. One model, one map, the low thinking level (which matters if you are paying per token), a harness the author wrote, and a mode that removes combat entirely. Those 13.5 hours are game time, not wall clock. It ranks Sol against nothing, and one run cannot say whether 13.5 hours is fast.

What it does supply is a failure mode you can watch. Madden's stated motivation is that Factorio is analogous to programming but far more legible to casual observers, which makes it a decent medium for building intuition about model capabilities. A production graph with a gap in it is easier to argue about than a benchmark score, and the argument is more specific: here is the belt that was never laid. He plans to test other models, other thinking levels, and runs with biters.