# GPT-5.6 Sol Beats Peaceful Factorio: What the Agent Run Actually Shows

A rocket launch after 13.5 hours of game time shows progress in long-running agent tasks, but the factory's unfinished connections matter too.

By Paras Madan | 2026-11-01


Ryan Madden reported on October 31 that GPT-5.6 Sol completed a peaceful Factorio run using Codex at a low thinking level. The agent reached a rocket launch in 13.5 hours of game time. It is a useful long-running agent demonstration, with important limits on what the result establishes.

This was not a new OpenAI model announcement or a standardized benchmark result. It was an independent experiment using a custom game interface, and peaceful mode removed enemy creatures from the task.

## How GPT-5.6 Sol played Factorio

Madden's [published account of the run](https://ryanmadden.net/gpt-56-beats-factorio/) describes a game manager and a small Python library through which the model controlled the game. The interface enforced restrictions such as the player's reach and fog of war, according to Madden.

That is different from a model operating the game entirely through screenshots, mouse clicks and keyboard input. A programmatic interface changes the work the agent has to do, even when it preserves game rules.

The distinction is useful rather than disqualifying. Many practical agents act through APIs and scripts. But an API-based success should not be presented as evidence that the same model can complete the task through any interface.

## A completed objective can hide an inefficient system

The run reached its goal, but Madden reports late-game scaling problems. Sol repeatedly carried supplies between factory areas to compensate for production systems it had failed to connect. The manual work in this description was done by the agent, not a claim that a human took over those steps.

Madden also reports improvement over his previous GPT-5.5 attempts, which struggled to use transport belts. That is a comparison within his experiments, not a controlled claim about the models across tasks.

The factory is a useful way to see a distinction that is harder to notice in a chat transcript. An agent can keep making progress by repeatedly repairing local problems while leaving the underlying system poorly organized.

In software work, the equivalent would be a task that passes its final check only because the agent repeatedly intervenes around a fragile process. Completion is valuable, but it does not tell an operator how much maintenance the result will need.

## Why Factorio is useful for studying agents

In his [earlier explanation of the project](https://ryanmadden.net/towards-automating-factorio/), Madden separates several challenges: spatial layout, chains of intermediate goals and production that happens asynchronously. He also distinguishes different levels of assistance, from prebuilt solutions to direct desktop interaction.

These details make a simple leaderboard comparison difficult. Giving an agent a better control library can change its result without changing the model. Removing enemies changes the planning problem. A long game also needs clear distinctions between simulation time, elapsed wall-clock time and model cost.

The headline's 13.5 hours refers to game time. The launch account does not provide enough information to calculate the cost of a successful run or a repeatable success rate across maps. It should not be used as a purchasing comparison between coding assistants.

## What a stronger follow-up would establish

The useful next evidence would be repeated runs under disclosed settings, including the full control interface, intervention rules, token costs and failure cases. It would also help to measure whether the factory remains productive without the agent repeatedly carrying materials around.

For builders evaluating long-running agents, that is the question worth carrying forward: after the agent finishes, has it built a system that keeps working, or become part of the machinery required to keep it running?

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain agent capabilities and their limits through clear technical writing and launch stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min)