Is Meta's new Muse Wonder model any good?

So Meta have just dropped their brand new family of models, Muse Wonder, and people have been saying some really nice things about it. Theo even put it in his S tier right next to Opus 5.5. But when I ran it through our Convex LLM leaderboard it only came out about the same as Haiku 5.5, so I wanted to give it a proper job and see for myself. So I got Opus 5.5 (in Claude Code), Muse Wonder and Muse Wonder Full (in Muse Code) to each build a mini version of Let's Build It, the vibe coding app for kids I recently launched with my son Josh, all from the exact same prompt. Then I have a look at what each of them built and grade them on their Convex code, the UI and how they cope when things go wrong. Have you tried Muse Wonder yet? Is it S tier for you? Let me know down in the comments. Links Convex LLM leaderboard: https://convex.dev/llm-leaderboard Let's Build It: https://letsbuildit.app Convex AI Gateway: https://docs.convex.dev/ai-gateway/overview Convex Auth v2 (alpha): https://auth-v2.previews.convex.dev/ Chapters 0:00 Is Muse Wonder really S tier? 1:27 The test: a mini Let's Build It 3:16 What they built 6:08 Mike from the future: em dashes and parent pages 10:46 Grading them 13:44 The verdict