GPT-5.6 Sol can beat Factorio on peaceful mode
Software engineer Ryan Madden published the write-up on his blog, ryanmadden.net, on October 31, 2026. He ran OpenAI's GPT-5.6 Sol through Codex at a low thinking level on Factorio's peaceful setting, where no biters attack, and the run reached a rocket launch after 13.5 hours of game time.
Madden used a game manager and a small Python library so the model could act through code rather than a mouse. The library enforces the same limits a human GUI player faces, including player reach and fog of war. He had planned to stop after blue science research, but Sol's performance led him to keep the run going.
The results were uneven. Scaling problems that Madden expected during the blue science stage arrived later and slowed Sol's progress. Sol did not connect some late-game production systems and instead carried supplies between factory areas by hand, which got it to the rocket. According to Madden's figures, relayed by RuntimeWire, the run used about 250 million input tokens, 95% of them cached, plus 1.5 million output tokens, which would cost roughly $52 at API prices.
Why it matters
For teams building agents, the run is a long-horizon test. Factorio demands hours of planning, construction and repair, so finishing a rocket launch shows a model can hold a plan together over many steps. The manual supply runs show the opposite risk: a model can reach a goal while hiding weak planning underneath. Teams that judge agents only by final completion will miss that failure mode.
For investors and operators, the cost figure matters more than the game result. A full run at about $52 in list API prices, with most input tokens served from cache, suggests that cache hit rates can shape the economics of long agent sessions as much as per-token price does. The result also comes from one author, one harness and one reasoning setting, so it is a useful signal for due diligence, not a benchmark claim.
Key points
- 1The run used GPT-5.6 Sol in Codex at a low thinking level and finished peaceful Factorio in 13.5 hours of game time.
- 2Peaceful mode removes biter attacks, so the test centers on factory planning, logistics and production.
- 3A Python library enforces human-like limits, including player reach and fog of war, on the model's actions.
- 4Sol failed to connect some late-game production systems and hand-carried supplies between factory areas to reach the rocket.
- 5
