Bowen compares the models through projects that require planning, tool use, iteration and physical or interactive verification rather than benchmark questions. In the 3D-design test, both models attempt to recreate a missing computer faceplate from measurements and photos. Soul produces the closer initial shape, but both require manual CAD changes, neither designs usable mounting clips, and the printed parts expose major limits in translating visual references into fabrication-ready geometry.
The creative-publishing task asks each model to commission, edit, illustrate and lay out a multi-author magazine using other models through an API. Fable assembles a coherent technology magazine with a clear editorial theme and documents the factual and continuity fixes it made. Soul produces a more visually ambitious but harder-to-interpret issue. Both results show that an agent can coordinate many specialist models, yet editorial coherence still depends on the choices made by the supervising system.
On a vintage iBook, both agents build and remotely deploy a simple 3D crime game. Soul packages the application with fewer dependencies and reaches a playable result quickly after repairing an old-processor instruction fault. Fable needs a longer tool-chain installation but delivers stronger graphics and uses the earlier crash log to avoid one failure. The test rewards not only the final game, but also autonomous diagnosis across obsolete hardware, networking and compilation constraints.
The mixed-reality and larger game tests split again. Soul's Vision Pro skateboard game is playable but awkwardly scaled, while Fable's repaired version feels more immersive. In the final narrative game, Soul creates technically impressive scenes but frustrating puzzle flow; Fable delivers a more complete cinematic story despite invisible characters, navigation bugs and other rough edges. Bowen's evidence suggests task-specific tradeoffs rather than a single model hierarchy: speed, packaging, visual quality, interaction design and recovery behavior can point to different winners. Subscription prompts and unrelated asides are omitted.
Watch on YouTube
