The same challenge
The extended input adds terrain, themed areas, dark ride stories, specialty rides, native voices, audio and a nighttime show. Its October 8 runs remain in the existing benchmark.
Read the original promptRead the extended prompt
ABOUT THE BENCHMARK
One benchmark.
Many park builds.
Original and extended prompt runs share the Theme Park benchmark, blind voting pool, Overall scoring and results graph. Each run identifies its published prompt version, exact input digest, model, effort, harness and available context.
Read the extended prompt
The original run
Play opens the submitted park. Product adaptations remain separate and do not change scored runs. Details show measurements and caveats; unrecorded effort is never inferred.
Review capture standardsBlind comparisons
Visitors compare captures in the same content category without model labels. Extra categories use the existing ranking method and activate with two eligible runs. They do not change the established Overall calculation. Native audio ballots play one option at a time; visitors can choose without listening.
Cast a blind voteComparing GPT-6.1 Sol runs
GPT-6.1 Sol names identify effort, CLI or Codex execution, recorded context size, and the extended prompt where used. Codex denotes the Codex desktop app; CLI denotes the Codex command-line launcher. K means 1,000 tokens. Known sizes use native usable context; the catalog or configured window is recorded separately in run details. Unavailable historical context is labeled explicitly.
The extended Max CLI and Desktop runs both used 258.4K usable context. The Desktop Ultra run used 828.4K. These runs help compare complete execution setups. Ultra changed both effort and context, so its result alone cannot isolate either effect. Repeated runs that change one setting at a time are needed to answer whether harness, context, or effort causes a difference.