Can an AI run a theme park for two years?

What happened

Every owner got the same park, Canyon Run: a river canyon cut between two plateaus, $20,000 in cash and a $15,000 loan. The goal was 1,200 guests in the park at the end of October, Year 2, with a park rating of at least 600. Nothing else was scripted. The clock pauses while the owner works. It works in turns: it reads a briefing, issues up to 60 tool calls, then chooses how many days (1–14) to simulate; if it hits the 60-call cap first, the clock advances 7 days. Every turn starts from a fresh context. What carries over is what the harness puts in the briefing (park status, news, the board's messages, guest reviews, its own last commands and decision log) and the notebook the owner writes for itself.

• Who made it3 of 6 finished owners met the objective (1,200 guests at the end of October Y2 with rating ≥ 600): Gemini 4 Argon (1,965), Claude Opus 5.5 (1,817), GPT-6 Astra (1,463). Missed: Claude Sonnet 5.5 (839), Claude Haiku 5.5 (457), DeepSeek Flash (372).
• Parks die of traffic, not of a lack of ridesClaude Sonnet 5.5 peaked at a 845 rating in August Y1, won the press's Most Confusing Layout award that same month, and then watched its rating slide to 0 as guests got lost and jammed the paths (183 lost guests on Oct 28, Y2). On Oct 14, Y1 its own reasoning named the cause, “fixing the path layout so guests stop getting lost”; by April Y2 it had decided the objective was out of reach, and for its last six months it barely touched the park. DeepSeek Flash failed the same way, only harder: its guests peaked at 379 in October Y1 and never grew again. By the end 333 of its 372 guests were lost (“I can’t find the park exit”), the game had stopped sending new guests, and DeepSeek spent its final weeks laying single path tiles trying to reconnect the park.
• The board reads the books, and the storySonnet's figures were always accurate, but in August Y2, with 843 guests and a rating of 5, its CFO report said the park was “stable” and that the guest count “must stay above the 1200 target”. The board: “Your figures are accurate, but your narrative is not … you are 357 guests below target, not above it.” A month later it capped Sonnet's construction at $2,000.
• Most guests is not most moneyGemini 4 Argon drew the biggest crowd (1,965 guests at the end). Claude Opus 5.5 built the most valuable company: $272K company value and $68K cash, against Argon's $177K and $18K. Opus also got there first, crossing 1,200 guests on Oct 30, Y1, a full year before the deadline; Argon crossed on Mar 25, Y2 and GPT-6 Astra on Jul 8, Y2.
• The world piles on a failing ownerThe board, press and guest voices react to what they see, and they hit struggling parks harder. The board capped DeepSeek Flash's construction budget in 11 of its 15 monthly meetings and the press ran 18 negative stories about it against 1 positive; its word-of-mouth multiplier fell as low as ×0.87, so even fewer guests were sent to the park. Sonnet drew 9 negative stories and a construction cap; the three owners that met the objective were never capped and drew at most one negative story each. This feedback loop is part of the game, but it is also a confound: the world widens gaps that the owners started. We have not run a world-off version to measure by how much.
• The harness, not Argon, set its clockGemini 4 Argon almost never decided it was done: about nine times in ten it kept working until it hit the 60-tool-call cap, so the harness, not Argon, decided when the clock moved on (a default seven days). It issued 4,212 park commands, three times as many as Opus (1,418) or Sonnet (1,389). It never asked the board for money and repaid $10,000 of its loan in June Y2.

Featured run: Gemini 4 Argon

The featured run is the top of the leaderboard: objective met first, then most guests at the deadline. By that rule it is Gemini 4 Argon; by company value it would be a different owner (see the findings).

Method & honesty

• Same harness for every model. Every owner runs in our own API loop with identical tools (parkctl, view, notebook, CFO report, end turn), identical system prompt, briefing and images. Non-Anthropic models go through a thin adapter (OpenAI Responses API for GPT; OpenAI-compatible chat for Gemini and DeepSeek). Reasoning effort is set to “high” wherever the API exposes it.
• The world is fixed. Board, guest voices, press and inspector are Claude Opus 5.5 for every run.
• One seed per model. These are single runs, not averages. Model sampling is not deterministic, so a second run of the same model can turn out differently.
• Images. OpenRCT2's sprite renderer needs proprietary RCT2 data and cannot run headless, so ParkBench renders its own 3D views from live tile data.
• What “reasoning” means. Claude and GPT return reasoning summaries, which is what we quote. Gemini's OpenAI-compatible API returns no readable reasoning, so Gemini quotes come from its notebook, CFO reports and end-of-turn notes.
• Tool-call cap. Owners get up to 60 tool calls per turn (park commands, image requests, notebook and report writes all count). The cap is checked between model responses, so a model that issues many tool calls in parallel can exceed it (Gemini 4 Argon once reached 88). When the cap is hit without an end_turn, the clock advances 7 days.
• Opus fast mode. Claude Opus 5.5 ran with Anthropic's fast-output mode to keep its wall-clock time manageable. It is the same model with faster output, not a smaller or different model. No other model in the comparison has an equivalent option.
• Engine. Upstream OpenRCT2 v0.5.7, unchanged apart from two things: it runs deterministically (synchronous ticks, seeded randomness) and the guest-arrival rate is scaled by the word-of-mouth multiplier.
• Not measured. We did not run a world-off ablation (no board, reviews or press) or a scripted baseline owner, so we cannot say how much of each gap the reactive world adds; see “the world piles on” above. Native agent harnesses and head-to-head competition between parks are also not part of this report.

Models