By Hamza Faisal
OpenAI's newest model has just cleared one of gaming's most demanding management challenges, mostly on its own. GPT-5.6 Sol, running inside the company's Codex coding environment, completed a full peaceful run of the factory-building game Factorio, launching a rocket after 13.5 hours of in-game time, according to the developer who set up the experiment.
Factorio asks players to start with almost nothing and build an ever-expanding automated factory: mining raw materials, smelting them, researching new technologies and finally assembling and launching a rocket into space. Launching that rocket is how you win the game, and doing it well requires planning dozens of interlocking production lines at once.
A factory run on autopilot
Developer Ryan Madden, who documented the run on October 31, gave Sol a low "thinking" setting inside Codex and connected it to the game through a custom Python game manager. The manager let the model play programmatically while enforcing the same limits a human player faces when using the game's interface, including restricted reach and fog of war, so the model could only act on what its in-game character could see and touch.
Sol handled the entire production chain, from early resource gathering through successive science milestones. Madden had originally planned to stop the run after blue science research, but Sol's early factory layouts were effective enough that he let it keep going all the way to the rocket launch.
Where the machine stumbled
The run was far from flawless, and the flaws are the interesting part. Scaling problems appeared in the late game, when Sol failed to properly connect some production systems. Rather than fixing the broken links, it papered over them, repeatedly ferrying supplies between disconnected factory sections by hand until the rocket was ready.
Even so, the result marks a clear step up from the previous generation. An earlier attempt with GPT-5.5 failed to use transport belts at all, while Sol built belts and organized its factories. The run also came with a price tag: Madden reported it used about 250 million input tokens, 95 percent of them cached, plus 1.5 million output tokens, working out to roughly $52 at API prices.
Why a game result matters beyond gaming
Factorio has become an unlikely benchmark for AI agents because it rewards exactly the skills that matter in real work: long-horizon planning, managing throughput, and recovering from mistakes over hours rather than minutes. A chat benchmark can be passed in seconds; a factory has to keep running.
There are fair caveats. The run used peaceful mode, so no enemy creatures attacked the factory, and the model had access to the vast library of Factorio guides written by human players. Madden has recorded the full run, released a timelapse, and plans to test other models, different thinking levels and non-peaceful runs next. As a legible test of what today's agents can sustain, though, a 13.5-hour rocket launch is hard to ignore.
