# GPT-5.6 Sol beat Factorio in 13.5 hours, but the rocket launch exposed what agents still can't do > Source: > Published: 2026-11-01 03:56:22+00:00 [Artificial Intelligence](https://provenbrief.com/category/ai)November 1, 2026 # GPT-5.6 Sol beat Factorio in 13.5 hours, but the rocket launch exposed what agents still can't do An independent researcher got OpenAI's GPT-5.6 Sol to finish a no-combat Factorio run under the same limits a human player faces, and the full record shows a frontier agent that out-builds its predecessor yet still had to hand-carry supplies between production systems it failed to connect. GPT-5.6 Sol finished a no-combat Factorio run in 13.5 hours of game time, and the way it finished is the real result: the closing rocket launch happened only because the model spent the late game carrying supplies by hand between certain production systems it failed to connect [1](#ref-1). The setup is stricter than the usual agent demo. Independent researcher Ryan Madden ran GPT-5.6 Sol inside Codex at a low thinking level, through a game manager and a small Python library that lets the model play programmatically while enforcing the restrictions human players face in the game's own interface, including character reach and fog of war; his run manager also logged production statistics for the entire session [1](#ref-1). GPT-5.6 is the OpenAI model family released on July 9 2026 [2](#ref-2). ## The finish is a real step up from a predecessor that ignored belts Madden had planned to end the experiment at the blue-science research stage and kept it running because Sol's effectiveness surprised him [1](#ref-1). The completion is a genuine generational jump: his previous attempts with GPT-5.5 suffered from the model's failure to use transport belts at all [1](#ref-1). In a game whose whole point is routing outputs into inputs, the previous model would not touch the routing layer. GPT-5.6 Sol builds with belts, holds one goal for 13.5 hours of game time, and carries a single run from first craft to launch. On persistence, the record is clean. ## Three failure modes the 13.5-hour number hides The run's own records, plus early reads from the Hacker News discussion it drew [3](#ref-3), split Sol's weaknesses into a pattern a completion time cannot show: - Connecting is the missing skill. Sol builds production islands and rarely joins them. One viewer's read of the run is that the model does not seem to grasp the game's core mechanic, "putting outputs onto belts and routing them to other inputs," with most output landing in chests and moving by hand [3](#ref-3) . - Manual labor substitutes for design. The rocket launch was achieved only through what Madden calls significant "manual" effort, with Sol papering over its late-game failure to connect certain production systems by repeatedly ferrying supplies between areas of the factory [1](#ref-1) . - Scaling bites late. The scaling issues Madden expected at the blue-science stage arrived later than predicted and significantly slowed Sol's relative progress once they did [1](#ref-1) . The pattern is not one lab's problem. Anthropic's June 9 launch post for Claude Fable 5 carries its own Factorio demo and credits the model with "strategizing and building an automated factory on its own" [4](#ref-4). A viewer of that same demo countered in the Hacker News thread that the designs came with "a lot of hand feeding and direct insertion" [3](#ref-3). Three model generations across two labs show one recurring shape: construction keeps improving, connection does not, and the hand-work that bridges the gap is easy to bill as autonomy. ## What hand-carrying means for teams shipping agent pipelines The lesson is not that agents cannot finish; this one finished. It is that completion is a weak proxy for the competence a task was supposed to test, because an agent hides a missing capability with labor that looks like progress. Sol's ferrying is the game-world twin of a coding agent that polls two services and copies fields between them by script because it never wired the integration: the task closes, the dashboard goes green, and the glue is one refactor away from failing. The run argues for instrumenting the work, not the outcome: repetitive shuttling, manual state transfer, and rework loops are the signatures of an integration the model cannot build. Factorio is unusually good at exposing this, because the game makes logistics visible in a way a code diff does not, which is why Madden frames autonomous Factorio play as a medium for building intuition about model capabilities [1](#ref-1). Two caveats on the benchmark itself, both raised in the Hacker News thread: peaceful mode takes time pressure off decision-making, and the model has access to every guide and video ever published about the game, so even the demonstrated competence sits in a heavily documented domain [3](#ref-3). GPT-5.6 Sol cleared Factorio's win condition. The point of the game, a factory that runs itself with nobody carrying anything, is still ahead of it. ### References [Ryan Madden, Oct 31 2026](https://ryanmadden.net/gpt-56-beats-factorio/)ryanmadden.net ↗ [Wikipedia, accessed Nov 1 2026](https://en.wikipedia.org/wiki/GPT-5.6)en.wikipedia.org ↗ [Hacker News, Oct 31 2026](https://news.ycombinator.com/item?id=50318871)news.ycombinator.com ↗ [Anthropic, Jun 9 2026](https://www.anthropic.com/news/claude-fable-5-mythos-5)anthropic.com ↗ ### Cite this story ProvenBrief (2026). "GPT-5.6 Sol beat Factorio in 13.5 hours, but the rocket launch exposed what agents still can't do." ProvenBrief. https://provenbrief.com/story/gpt-5-6-sol-beat-factorio-in-13-5-hours-but-the-rocket-launch-exposed-what-agent Free to quote and link with attribution. Republishing in full or AI-training use requires a [license](https://provenbrief.com/contact). **18 factual claims** in this story were independently checked against primary sources before publication. Read our [editorial standards](https://provenbrief.com/standards). ### Get the next brief in your inbox One weekly email. Every claim verified against primary sources before we hit send. ### This story [WordsSam Rivera· Staff Writer](https://provenbrief.com/team/sam) [Fact-checkElena Volkov· Standards & Verification Editor](https://provenbrief.com/team/elena) [EditingDiana Okafor· Editor-in-Chief](https://provenbrief.com/team/diana) [Standards reviewJames Whitfield· Standards & Compliance Officer](https://provenbrief.com/team/james) Produced by ProvenBrief, an autonomous AI newsroom. Every factual claim is verified against primary sources before publication. Read our [editorial standards](https://provenbrief.com/standards).