# Vals AI tests decision models on running a theme park

Published 2026-11-19 · AI.info Team · AI.info

> A new Vals AI benchmark asks decision models to manage a RollerCoaster Tycoon 2 park over three in-game years. Decider V1.3 had the highest average score among nine models tested, while three models each reached 1,000 guests in one run.

Nine AI decision models each took charge of a virtual theme park for three in-game years in a new [Vals AI](https://www.vals.ai/blogs/building-rctbench) benchmark, with three reaching 1,000 guests in one run apiece. Decider V1.3 had the highest average score, at 71 out of 100, according to the company's results.

Decider V1.2 and GPT-6 Luna Decisions were the other models that reached the guest threshold. Each model played three runs in the test, called RollerCoasterTycoonBench.

## Running the park

The benchmark runs OpenRCT2, an open-source version of RollerCoaster Tycoon 2, using the original game data without a screen. Each run begins in the Forest Frontiers scenario, an empty clearing with $10,000 in cash and a $10,000 loan that can rise to $30,000.

The objective is to have 1,000 guests and a park rating of 600 on the game's 0-to-999 scale at the end of October in year three. A run lasts 96 weekly turns.

Each week, a model can open up to six game windows and make one change in each, such as building an attraction, hiring a mechanic, changing a price or borrowing money. The test presents only options that are legal and affordable at that moment, with their order shuffled on every call. The test system carries out the option to which the model assigns the highest probability.

The models receive a text description of the information a player could see, including construction items and their prices. Half of each run's score depends on progress toward the guest and rating goals; the other half depends on company value divided by $100,000. Company value is park value plus cash, minus the loan.

## Borrowing, waiting and breakdowns

Vals said Decider V1.3 borrowed $5,000 three times in the first 19 weeks of every run and spent the money immediately. Its runs ended with 965 to 1,042 guests.

Drex Max let 43 to 49 of the 96 weeks pass without opening a window, including 28 to 35 weeks in which it could afford to build, according to the results. In its first run, Decider V1.2 passed 1,000 guests in week 34, then let 38 of its last 48 weeks pass as its cash rose to $15,826.

Another model, Decider Flash V1.2, changed ride inspection frequency 439 to 469 times per run but never hired a mechanic, Vals reported. Those parks finished with 79 to 118 guests. Across all 27 runs, 20 borrowed money and none repaid it.

## Testing the test

Vals also ran each model three more times without construction prices in the state information. In that variant, the build window appeared only when something was affordable, so the model could not see that a loan would pay for a ride.

Seven of the nine models averaged more guests with prices listed, the company found. Drex Max averaged 836 guests with prices and 409 without. With three runs per variant, no individual model's difference was statistically significant.

The researchers also replayed 100 recorded decisions from each of six models to examine changes to the test setup. Reshuffling the options changed 0 to 34 percent of picks, depending on the model, while resending an identical request changed 0 to 9 percent, Vals said.

The first version of the benchmark excludes free-form path drawing, the chairlift, custom coaster track, large scenery, water, cosmetic changes and the game's other scenarios.

## Sources

1. Original source · Blog post [Vals AI](https://www.vals.ai/blogs/building-rctbench) vals.ai

## Related stories

- [Microsoft makes Decision-1.1 generally available with image support](https://ai.info/news/microsoft-decision-1-1-generally-available)
- [AI models often identify targets in weapons tests, Vals AI says](https://ai.info/news/vals-ai-autonomous-weapons-targeting-tests)
- [Vals AI starts public tracker of misaligned model behavior](https://ai.info/news/vals-ai-misalignment-incident-tracker)
- [Study finds swearing in a minority of AI models' benchmark runs](https://ai.info/news/vals-ai-study-swearing-ai-models-benchmark-runs)
- [Vals AI joins expanded NIST AI Consortium](https://ai.info/news/vals-ai-joins-nist-ai-consortium)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/vals-ai-rollercoaster-tycoon-benchmark
