Bijan Bowen tests both Muse Wonder variants across a browser desktop, skate and subway games, a pool-party game, a watch website and a robot-arm task. Initial outputs range from competent details to simple visuals and functional omissions, so he does not treat launch benchmarks as proof of one-shot polish.
Bijan Bowen gets a much stronger skate-game result after giving Wonder Full a reference image in a fresh session and allowing extended iteration. A subway-game follow-up adds detail without a reference, whereas the non-full model's robot-arm attempt remains unsuccessful despite cautious movement. The tasks and variants differ, so these are practical observations rather than a controlled overall ranking.
Bijan Bowen initially reports severe quota consumption but later observes resets and describes a usage bug, withdrawing a firm judgment about subscription value. He sees signs of fast, efficient work from the API-accessible variant while retaining concerns about price and the gap between benchmark expectations and visible results.
Watch on YouTube
