AI Code King compares five model and harness combinations across six one-hour projects using fixed prompts and a common scoring rubric. Interrupted outputs remain in the assessment, while Haiku's later run uses a newer harness. The results describe a small set of workloads, not a controlled demonstration that one model is universally superior.
AI Code King follows features through actual use: emulator support remains partial, container replacements do not prove efficiency or isolation, and room-editor saving can fail. Other tests distinguish local keyboard editors from working remapping, winning a game from preserving progress, and processing synthetic fitness data from verified live-device synchronization.
AI Code King estimates API-equivalent costs for the three projects with complete usage records. Those estimates do not represent subscription charges or all six projects. The broader lesson is to trace the feature to its final outcome and inspect failure boundaries rather than treating generated code, screenshots or a single score as sufficient evidence.
Watch on YouTube
