GPT-5.6 Sol Beats Factorio: OpenAI's Model Launches the Rocket

On October 31, Ryan Madden described on his blog how GPT-5.6 beats Factorio: GPT-5.6 Sol, running in Codex at low reasoning effort, finished the game on peaceful mode, meaning no enemies. From the start to the rocket launch took 13.5 hours of game time. The post hit number one on Hacker News.

How did GPT-5.6 Sol play Factorio?

Factorio is a factory-building game: you go from mining ore by hand, through conveyor belts and assemblers, to ever more advanced science packs and finally a rocket sent into space. Launching the rocket is the classic end of the game.

The model never looked at the screen. It played through code: a game harness gave it a small Python library to control the character. Crucially, the library enforced the same limits as the human interface, including the character’s reach and the fog of war. So the model couldn’t grab objects on the other side of the map.

According to the author, Sol exceeded his expectations in some areas and showed clear weaknesses in others. Madden had planned to stop the test after blue science, but the model’s performance convinced him to keep going. He recorded the whole run, and the timelapse is on YouTube.

Where did Sol struggle?

The ending is the most interesting part. The scaling problems the author expected around blue science arrived later and slowed progress badly. The rocket only launched because Sol patched disconnected production lines by carrying supplies by hand between parts of the factory. The author calls this the model’s “manual” effort (in quotes, because it’s still a model, not a person).

One commenter points out the same thing: instead of putting products on belts and routing them to the next machines, the model often dumped everything into chests and carried it around with the character. Another summed it up: “Robots: they’re just like us!” ;)

What’s new here for AI?

The progress is clearest against the same author’s earlier attempts:

1. In March, Madden wrote in “Towards beating Factorio autonomously” that the first autonomous rocket launch was “still up for grabs”. The best results from the Factorio Learning Environment stopped at logistics science packs in the full game.
2. In April, he and a friend hooked Claude Code up to Factorio. Claude built a working power plant and production of red and green science, and they ended the test halfway through the “automation 2” research. That tool, however, let the model reach objects anywhere on the map and mine without delay.
3. Now GPT-5.6 Sol has reached the rocket, under the same limits a player has. The author also highlights a clear improvement over GPT-5.5, which couldn’t use conveyor belts at all.

It’s a bit like the recent case of Gemini in Minecraft, where agents also played by writing code. Madden thinks Factorio resembles programming but is much easier for a viewer to follow, which makes it a good way to show what models can really do.

What should you keep in mind?

First, the game mode. This was a peaceful run, without the enemies that attack a growing factory in a normal game. The author plans attempts with enemies.

Second, the scale of the test. This is one run of one model, described by one person. Madden didn’t say how long the run took in real time or what it cost, and the harness code isn’t linked in the post.

Third, the quality of the factory. The rocket launched, but the factory was far from what an experienced player builds. A lot depended on carrying supplies by hand.

The author promises a commentated video of the whole run and tests of other models and reasoning levels. If you want to know where GPT-5.6 Sol sits in the rest of OpenAI’s lineup, I covered that in the GPT-6.1 Terra launch.

FAQ

Did GPT-5.6 Sol beat Factorio on its own?

According to the author, yes, on peaceful: the model controlled the game through a Python library and got to the rocket launch after 13.5 hours of game time. The post doesn’t mention any human help during the run.

What is peaceful mode in Factorio?

It’s a world setting without enemies (the “biters”) that normally attack your factory. The game is then only about building and logistics.

How much did the experiment cost?

We don’t know. The author didn’t give the cost or the number of tokens, and on Hacker News it’s one of the first questions people asked.

Can I try it myself?

The tool used for this run hasn’t been released yet. The author previously published Faketorio, a simple agent sandbox inspired by the early game of Factorio.