What happened in the Factorio run?
On October 31, developer Ryan Madden reported a completed Factorio run using GPT-5.6 Sol inside Codex at low thinking level. The agent reached the rocket launch in 13.5 hours of game time on peaceful mode, with no biters. Madden published a full-run timelapse.
A game manager connected Codex to a small Python library. Madden says the interface enforced human-player restrictions including character reach and fog of war. This was programmatic interaction, not a demonstration of visual mouse-and-keyboard control.
The result is notable because reaching a rocket requires coordinating resource extraction, manufacturing, power and research over many steps. It does not establish a general benchmark win or a repeatable success rate.
| Detail | What the source reports |
|---|---|
| Model and harness | GPT-5.6 Sol in Codex |
| Reasoning setting | Low thinking |
| Environment | Peaceful mode; no biters |
| Completion | Rocket launch after 13.5 game hours |
| Interaction | Python library through a game manager |
| Total bill | Not disclosed in the post |
Where did Sol struggle?
Madden originally planned to stop after blue science research, but continued because the agent was doing better than expected. Scaling problems emerged later and substantially slowed progress.
The important caveat is how it finished: Sol failed to connect some production systems and repeatedly carried supplies between factory areas. The source calls this significant “manual” effort by Sol, not a claim that a human completed those steps.
That distinction matters for agent buyers. Reaching an objective and building a robust automated system are different acceptance criteria. An agent can compensate for a faulty design with repeated actions while accumulating latency and usage.
Madden says earlier GPT-5.5 attempts failed to use transport belts at all. That is an anecdotal comparison across attempts, not a controlled model ranking. The post does not provide repeated-trial success rates or a matched cost comparison.
What does this mean for pricing?
No new rate card, price cut or plan entitlement is announced here. The source does not disclose the run’s token totals, billing route or total cost. Game time is not inference time, wall-clock duration or billable usage.
The generated table, chart and calculator use the maintained canonical data, with separate per-model price dates. This release reconciled every retained Sol and successor-reference billing dimension against the current official pricing documentation, including cache writes, Standard and Batch prices, and the full-request long-context tier. Rates did not change. Merely fetching a source, updating this article or building the site cannot verify a price.
OpenAI’s Sol model documentation says promotional pricing is available at least through November 21, 2026; no post-promotion rate is established here. A Codex demonstration does not establish a particular ChatGPT plan entitlement. Asynchronous Batch rates are not an assumed discount for the interactive game loop.
GPT-5.6 Sol is classified as legacy in our catalog. GPT-6.1 Sol is shown only as a successor-rate reference: Madden did not test it in this report. Neither the chart nor the calculator predicts Factorio performance or a Codex subscription bill.
For the full rate card, use OpenAI pricing. Compare a separate provider through Anthropic pricing, without assuming this run establishes Claude’s relative performance.
Labs: included text route, Factorio replay blocked
Our published Cost-per-Task Labs dataset contains the exact openai/gpt-5.6-sol route. That dated deterministic text test is not a Factorio result, a current account-access probe or a reproduction of Madden’s run.
The Factorio replay blocker requires a pinned game build and map seed, the Python library and game-manager version, exact Codex/model settings, action restrictions, starting state, stopping rules, complete trajectories, an intervention log, repeated trials, independently checked rocket completion, and a full token/cache/tool/compute billing ledger. No paid experiment was run for this report.
What should agent builders do now?
Builders evaluating long-running agents gain a useful failure case to study. Budget owners still lack the evidence needed to call this run economical; the report establishes no provider winner or loser on cost.
If you are already using Sol, test low thinking on a representative task before increasing reasoning effort. This demonstration shows one success at that setting, not a guarantee that low thinking is best for your workload.
Set an explicit completion test, a tool-call ceiling and a spend limit. Record token usage, retries, elapsed time and human interventions. Inspect whether the agent solved the underlying workflow or merely kept compensating for broken connections.
Use the token calculator for hypothetical API workloads and our API cost calculation guide to separate rates from total task cost. Do not convert 13.5 game hours into a bill without usage logs.
Bottom line: this is an interesting long-horizon agent demonstration with visible engineering weaknesses. It is not evidence of a new model release, autonomous success against enemies, or a known cost per completed game.