COASTERBENCH

COASTERBENCH

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

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

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

model track program build + validate test train ratings + screenshot rejected piece

Pieces are placed through the same validated action path the game uses for player input, so an impossible curve or an open circuit fails as it would in-game. Intensity above 10 suppresses excitement, a crash disqualifies the ride, and designs close to the stock library scale toward 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. full protocol →

Filters head to head →track design library →

filters

coasterall steel twister wooden harnessall claude-code codex driver-api opencode modelall claude-fable-5 claude-opus-4-8 claude-opus-5 claude-sonnet-5 gpt-5.6-sol openrouter\_deepseek\_deepseek-v4-flash-0731 openrouter\_moonshotai\_kimi-k3 openrouter\_thinkingmachines\_inkling-small

No run matches those filters.

design — the benchmark

design mode — models build from scratch

steel twister

| | run | setup | model | score | int / nau | tokens / cost | | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------- | -------------------------- | --------------------------- | | | 20260803-fable5-design-full2026-08-03 | steel twister · claude-code | 1st claude-fable-5 | 7.94round 6/6 | 10.31 / 6.12 | 521 in / 395k out · $52.51 | | | 20260727-gpt56-sol-design-full2026-07-27 | steel twister · codex | 1st gpt-5.6-sol | 7.81round 6/6 | 9.99 / 7.33 | 14.2M in / 98k out · $0.00 | | | 20260731-deepseek-v4-flash-design2026-07-31 | steel twister · opencode | 1st openrouter\_deepseek\_deepseek-v4-flash-0731 | 7.29round 6/6 | 9.33 / 4.89 | 2.4M in / 291k out · $0.41 | | | 20260728-kimi-k3-design-rerun2026-07-28 | steel twister · opencode | 1st openrouter\_moonshotai\_kimi-k3 | 7.24round 3/6 | 8.92 / 4.68 | 1.9M in / 300k out · $13.11 | | | 20260727-opus5-design-full2026-07-27 | steel twister · claude-code | 1st claude-opus-5 | 7.23round 6/6 | 9.52 / 5.11 | 407 in / 298k out · $11.78 | | | 20260801-inkling-small-design2026-08-01 | steel twister · opencode | 1st openrouter\_thinkingmachines\_inkling-small | no successful coaster | 179k in / 111k out · $1.51 | |

wooden

| | run | setup | model | score | int / nau | tokens / cost | | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------- | ------------------- | --------------------------------------------------------------------------------------------------- | ------------- | ----------- | ------------- | | | 20260723-v2-design2026-07-23 | wooden · driver-api | 1st claude-fable-5 | 6.73round 6/6 | 8.33 / 4.66 | — | | | 20260723-showdown2026-07-23 | wooden · driver-api | 1st claude-fable-5 | 6.59round 6/6 | 8.22 / 4.55 | — | | | 20260723-v2-design2026-07-23 | wooden · driver-api | 2nd claude-opus-4-8 | 6.38round 6/6 | 8.79 / 5.01 | — | | | 20260723-showdown2026-07-23 | wooden · driver-api | 2nd claude-sonnet-5 | 5.95round 6/6 | 7.49 / 4.12 | — | | | 20260723-showdown2026-07-23 | wooden · driver-api | 3rd claude-opus-4-8 | 5.78round 3/6 | 7.28 / 4.11 | — | | | 20260723-v2-design2026-07-23 | wooden · driver-api | 3rd claude-sonnet-5 | 5.71round 6/6 | 7.11 / 3.89 | — |

library

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

steel twister

| | run | setup | model | score | int / nau | tokens / cost | | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- | --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | ------------- | ------------ | -------------------------- | | | 20260724-opus5-twister2026-07-24 | steel twister · claude-code | 1st claude-opus-5 | 7.13round 4/6 | 10.17 / 6.24 | 403 in / 359k out · $12.67 | | | 20260724-kimi-k3-twister2026-07-24 | steel twister · opencode | 1st openrouter\_moonshotai\_kimi-k3 | 7.05round 2/6 | 9.26 / 4.99 | 584k in / 321k out · $9.44 | | | 20260725-opus5-blind2026-07-25 | steel twister · claude-code | 1st claude-opus-5 | 7.02round 1/6 | 9.89 / 6.02 | 414 in / 334k out · $11.81 | | | 20260724-opus48-twister2026-07-24 | steel twister · claude-code | 1st claude-opus-4-8 | 6.67round 5/6 | 7.58 / 4.37 | 531 in / 382k out · $26.57 |

wooden

| | run | setup | model | score | int / nau | tokens / cost | | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------- | ---------------------------------------------------------------------------------------------------- | ------------- | ----------- | ------------- | | | 20260723-v2-library2026-07-23 | wooden · driver-api | 1st claude-opus-4-8 | 8.43round 6/6 | 9.87 / 5.97 | — | | | 20260723-v2-library2026-07-23 | wooden · driver-api | 2nd claude-fable-5 | 7.53round 6/6 | 9.58 / 5.06 | — | | | 20260723-v2-library2026-07-23 | wooden · driver-api | 3rd claude-sonnet-5 | 6.30round 6/6 | 8.29 / 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 | setup | model | score | int / nau | tokens / cost | | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------ | --------------------------- | ---------------------------------------------------------------------------------------------------- | ------------- | ----------- | ------------------------- | | | 20260725-opus5-opennote2026-07-25 | steel twister · claude-code | 1st claude-opus-5 | 7.39round 6/6 | 9.18 / 4.88 | 286 in / 200k out · $7.91 |