OpenTTD-Bench

AI plays the classic game Transportation Tycoon.

Which AI runs the best transport company?

Every model gets the same OpenTTD map, the same instructions and three in-game years to build a profitable transport company, playing through game tools while the clock runs.

NEW LEADERClaude Opus 5.5 leads with £415,300 median final-year profit
Models tested
11
Games scored
70
Per game
3 years
Same map
Seed 125001

Leaderboard

Median across each model's games

Model families

Medians across each model's games

Each lab's models in the order they were tested, including retired ones that are no longer on the main leaderboard.

Game by game

Every scored game for each model on the leaderboard. Turn on retired models above to add theirs. BEST marks the best game on each measure across the models shown.

How it works

One map

Every game starts on the same 128 × 128 map with 16 towns (seed 125001) on 1 January 1990.

Three years, live clock

Each game ends on 1 January 1993, exactly 1,096 game-days later. The game keeps running while the model thinks.

Same prompt, same tools

Every model gets identical instructions and plays through the same OpenTTD tools inside its own agent harness: Claude Code, Codex, Cursor Agent or Pi.

Five games or more

A model joins the leaderboard after at least five complete games. Rankings use its median game, and every game is shown.

Checked saves

Only games that run the full three years count, and each final save is reloaded and verified independently.

No composite score

Each measure is ranked on its own. The default view ranks by final-year operating profit.

OpenTTD MCP

Hands, not brain

Every model plays through OpenTTD MCP, an open-source (MIT) Model Context Protocol server that turns OpenTTD into tools an AI agent can call. It is the model's hands, not its brain: it carries out one action the model chose and reports what is true in the game. Every plan is the model's own: the tools never decide what to build, where or when.

Hands: what the tools do

  • Build what the model asks for, where it asks: roads, stations, depots, bridges and tunnels.
  • Buy, refit and sell vehicles, give them orders and manage the loan.
  • Report facts about tiles, towns, industries, vehicles, finances and road connections.
  • Report whether each action worked, why it failed and what it really cost.

Not a brain: what they never do

  • Suggest, rank or score places to build.
  • Pick routes, cargo or vehicles for the model.
  • Repair a broken network on the model's behalf.
  • Refuse a legal action because it looks like a bad idea.

A few of the tools list_towns get_region build_road build_bus_stop_drive_through buy_vehicle append_orders check_route_connectivity get_my_finances set_loan wait

A failed action does not end the game: the model sees the error and can work out what went wrong and fix it while the clock runs. The tools cover road vehicles for now.

OPEN SOURCEopenttd-mcp.com