Bijan Bowen evaluates Claude Haiku 5.5 using a browser-based desktop, a dependency-free C++ skateboarding game and several game-development tasksGame generation uses AI to create or assemble playable assets, scenes, rules, audio and interactions.. The results include functional interfaces and controls alongside visual defects, missing interactionsCoding output quality measures whether AI-generated software is correct, maintainable, complete and suitable for its intended use. and an output-limit failure that requires a different reasoning settingReasoning effort is the amount of internal computational work an AI model applies before producing an answer or action..
Bijan Bowen revisits the skateboarding result with reference-image feedback and diagnoses a subway game's rendering and freezing problems through multiple follow-ups. A pool-party example demonstrates asset creation and water effects, while other tests expose simplified mechanics and inconsistent visual fidelity.
Bijan Bowen also tests a watch interface and a demolition-derby scene. The concluding assessment separates these mixed capabilities from Bijan Bowen's reported subscription-usage measurement: the tested workload consumes little of his weekly allowance, but that observation is not a general cost or performance guarantee.
Watch on YouTube
