AI beats Factorio: GPT-5.6 Sol launches a rocket, GPT-6 Astra does it far faster

What's going on

Developer Ryan Madden has shown that GPT-5.6 Sol, running in the Codex coding agent at a low thinking level, can play the factory-building game Factorio in peaceful mode (no hostile "biters") all the way to a rocket launch. The run took 13.5 hours of game time. A timelapse is on YouTube, and a commentary video is on the way. Two more reports (The Decoder, Brief IA) describe how an even newer model, GPT-6 Astra, masters several games far faster.

How the Factorio run was set up

As in Madden's earlier experiments (Claude Code plays Factorio, "Towards beating Factorio autonomously", the "Faketorio" test setup), the AI doesn't control the game with a mouse. It uses a "game manager". This time it was a small Python library; before, it was Bash. It enforces the same limits that human players face in the interface, such as the player character's reach and fog of war (unexplored areas are invisible). The run manager also recorded production statistics. Madden published a chart of total resource production with markers for completed research.

Sol's strengths and weaknesses

Madden writes that Sol exceeded his expectations in some areas and showed clear weaknesses in others. He originally planned to end the run after blue science, but let it continue because it was so effective. After that, the expected scaling problems appeared and slowed progress considerably.

Sol reached the rocket launch only with a lot of "manual" effort. In the late game it failed to connect certain production systems, so it compensated by repeatedly hauling materials back and forth between factory areas. Even so, Madden sees a clear improvement over earlier attempts with GPT-5.5, which couldn't use conveyor belts properly at all.

Madden thinks autonomous Factorio play is a good way to get a feel for model capabilities. It resembles programming, but is far easier for non-experts to follow. He plans to test more models, thinking levels and runs with enemies.

What's reported about GPT-6 Astra

According to The Decoder and Brief IA (which cite operators, Vals AI and ARC Prize), GPT-6 Astra clearly beats GPT-5.6 Sol:

Astra's Factorio run was a community run, driven through a custom Lua mod with an MCP interface (a standard that lets AI models call tools). It isn't directly comparable to Madden's setup, and Space Age is an expansion of the game. According to The Decoder, Astra also got through Portal to the credits, finished Fallout 2 in 22 hours, and led a RimWorld colony off the planet.

Learning the rules, and overcorrecting

ARC Prize explains the jump on ARC-AGI-3 (a benchmark of unfamiliar, abstract game worlds where you have to infer the rules by trial and error) like this: Astra translates unfamiliar mechanics into compact symbolic descriptions. It tracks objects, coordinates, rules and plans in a shorthand of its own devising. Observations become rules, and rules become plans.

In the Fallout 3 run (user imjustnewatai, manual start, then Codex at maximum thinking level; saving, companions and fast travel allowed), this showed up as a principle: when an input has no effect, Astra doesn't blindly repeat it but looks for the cause. For example, a bed at first didn't heal, but after the game ran on briefly, health was back at 260/260.

In Minecraft, the behavior tipped over. Vals AI had Astra play purely through screen, mouse and keyboard, with no special interface. In under three hours it built a Nether portal, built a semi-automatic blaze farm, and collected six blaze rods and three ender pearls. According to Vals AI, that's further than any AI has gotten. Then a creeper blew up its chest and bed. Astra noted a rule to always carry important items, started distrusting large green objects (it mistook sugar cane for a creeper), and spent hours in the potato field. Vals AI called it off after 141 hours.

Our take

Madden's result is a sober data point: a low-thinking-level model can complete a full Factorio playthrough, but only with workarounds and in peaceful mode. The Astra numbers come from community and vendor environments with different tools and can hardly be compared directly. Read the 99.9% with OpenAI's own harness with particular caution. One striking point is that learning from mistakes is both a strength and a risk: a bad experience turns into a permanent rule, even when it no longer helps.