Show HN: Having models make roller coasters instead of pelicans (via OpenRCT2)

Disp4tch1 pts0 comments

CoasterBench

COASTERBENCH<br>A benchmark in which language models design roller coasters that RollerCoaster Tycoon 2 builds, tests, and rates.

Sevenfold Serpent<br>claude-fable-5<br>excitement 7.94<br>intensity 10.31<br>nausea 6.12<br>length 1369 m<br>drops 21<br>airtime 139<br>pieces 211

Each round a model submits a track program. OpenRCT2 builds it, runs a test train, and<br>returns ratings and a screenshot; the reported score is excitement .

model<br>track program

build<br>+ validate

test train

ratings<br>+ screenshot

rejected piece

Pieces are placed through the same validated action path the game uses for player input, so<br>an impossible curve or an open circuit fails as it would in-game. Intensity above 10 suppresses<br>excitement, a crash disqualifies the ride, and designs close to the stock library scale toward<br>zero. Ratings come from the game, not from a model. Modes: design — design mode — models build from scratch. library — library mode — models may search the stock track design library, which measures retrieval and adaptation; copies score zero. open note — open-note condition — the agent is given the engine source it is scored by, so it can read the ratings code instead of guessing; scores are not comparable with black-box runs.<br>full protocol &rarr;

Filters<br>head to head &rarr;track design library &rarr;

filters

coaster<br>all<br>steel twister<br>wooden

harness<br>all<br>claude-code<br>codex<br>driver-api<br>opencode

model<br>all<br>claude-fable-5<br>claude-opus-4-8<br>claude-opus-5<br>claude-sonnet-5<br>gpt-5.6-sol<br>openrouter_deepseek_deepseek-v4-flash-0731<br>openrouter_moonshotai_kimi-k3<br>openrouter_thinkingmachines_inkling-small

No run matches those filters.

design — the benchmark

design mode — models build from scratch

steel twister

run<br>setup<br>model<br>score<br>int / nau<br>tokens / cost

20260803-fable5-design-full2026-08-03<br>steel twister &middot; claude-code<br>1st claude-fable-5<br>7.94round 6/6

10.31 /<br>6.12<br>521 in / 395k out · $52.51

20260727-gpt56-sol-design-full2026-07-27<br>steel twister &middot; codex<br>1st gpt-5.6-sol<br>7.81round 6/6

9.99 /<br>7.33<br>14.2M in / 98k out · $0.00

20260731-deepseek-v4-flash-design2026-07-31<br>steel twister &middot; opencode<br>1st openrouter_deepseek_deepseek-v4-flash-0731<br>7.29round 6/6

9.33 /<br>4.89<br>2.4M in / 291k out · $0.41

20260728-kimi-k3-design-rerun2026-07-28<br>steel twister &middot; opencode<br>1st openrouter_moonshotai_kimi-k3<br>7.24round 3/6

8.92 /<br>4.68<br>1.9M in / 300k out · $13.11

20260727-opus5-design-full2026-07-27<br>steel twister &middot; claude-code<br>1st claude-opus-5<br>7.23round 6/6

9.52 /<br>5.11<br>407 in / 298k out · $11.78

20260801-inkling-small-design2026-08-01<br>steel twister &middot; opencode<br>1st openrouter_thinkingmachines_inkling-small<br>no successful coaster<br>179k in / 111k out · $1.51

wooden

run<br>setup<br>model<br>score<br>int / nau<br>tokens / cost

20260723-v2-design2026-07-23<br>wooden &middot; driver-api<br>1st claude-fable-5<br>6.73round 6/6

8.33 /<br>4.66

20260723-showdown2026-07-23<br>wooden &middot; driver-api<br>1st claude-fable-5<br>6.59round 6/6

8.22 /<br>4.55

20260723-v2-design2026-07-23<br>wooden &middot; driver-api<br>2nd claude-opus-4-8<br>6.38round 6/6

8.79 /<br>5.01

20260723-showdown2026-07-23<br>wooden &middot; driver-api<br>2nd claude-sonnet-5<br>5.95round 6/6

7.49 /<br>4.12

20260723-showdown2026-07-23<br>wooden &middot; driver-api<br>3rd claude-opus-4-8<br>5.78round 3/6

7.28 /<br>4.11

20260723-v2-design2026-07-23<br>wooden &middot; driver-api<br>3rd claude-sonnet-5<br>5.71round 6/6

7.11 /<br>3.89

library

library mode — models may search the stock track design library, which measures retrieval and adaptation; copies score zero

steel twister

run<br>setup<br>model<br>score<br>int / nau<br>tokens / cost

20260724-opus5-twister2026-07-24<br>steel twister &middot; claude-code<br>1st claude-opus-5<br>7.13round 4/6

10.17 /<br>6.24<br>403 in / 359k out · $12.67

20260724-kimi-k3-twister2026-07-24<br>steel twister &middot; opencode<br>1st openrouter_moonshotai_kimi-k3<br>7.05round 2/6

9.26 /<br>4.99<br>584k in / 321k out · $9.44

20260725-opus5-blind2026-07-25<br>steel twister &middot; claude-code<br>1st claude-opus-5<br>7.02round 1/6

9.89 /<br>6.02<br>414 in / 334k out · $11.81

20260724-opus48-twister2026-07-24<br>steel twister &middot; claude-code<br>1st claude-opus-4-8<br>6.67round 5/6

7.58 /<br>4.37<br>531 in / 382k out · $26.57

wooden

run<br>setup<br>model<br>score<br>int / nau<br>tokens / cost

20260723-v2-library2026-07-23<br>wooden &middot; driver-api<br>1st claude-opus-4-8<br>8.43round 6/6

9.87 /<br>5.97

20260723-v2-library2026-07-23<br>wooden &middot; driver-api<br>2nd claude-fable-5<br>7.53round 6/6

9.58 /<br>5.06

20260723-v2-library2026-07-23<br>wooden &middot; driver-api<br>3rd claude-sonnet-5<br>6.30round 6/6

8.29 /<br>4.62

library + open note

open-note condition — the agent is given the engine source it is scored by, so it can read the ratings code instead of guessing; scores are not comparable with black-box runs

run<br>setup<br>model<br>score<br>int / nau<br>tokens / cost

20260725-opus5-opennote2026-07-25<br>steel twister &middot; claude-code<br>1st claude-opus-5<br>7.39round 6/6

9.18 /<br>4.88<br>286 in / 200k out · $7.91

claude middot steel twister design wooden

Related Articles