Muse Wonder ranks first on three agent tasks but produces weak design, Claire Vo's benchmark finds

Meta’s new Muse Wonder models are good at agent work and bad at design, according to a first-look benchmark from Claire Vo, host of the How I AI podcast and founder of ChatPRD. If you were hoping one model could both build and style an app, her results point the other way.

Strong on agent tasks, weak on design

Vo’s write-up for Lenny’s Newsletter is titled “Meta’s new ‘Muse Wonder’ model: good for AI agents, bad for design.” She says she ran Muse Wonder and Muse Wonderful “through the How I AI bench, which covers nearly 20 tasks across prototyping, agentic workflows, coding, writing, and creative work.” She scored the results with “an LLM-as-a-judge pass and my own voice notes.” In an LLM-as-a-judge pass, another AI model grades the outputs, so part of the scoring is a model’s opinion rather than a fixed test.

Her episode outline lists three findings. Muse ranked first on three agentic tasks, its design output is “(it’s not good),” and it is “fast on one-shot work, slow on agent tasks.” She also compares it with Sonnet on voice assistant tasks and with Opus 5.5, Astra, and Fable on SVG illustration and short-form video.

Fast on one request, slow on agent tasks

The speed finding matters most for anyone choosing a model for agent work. A model that answers a single request quickly can still make a long job drag, because every step in the job adds its own wait. Judge it by the time a whole task takes, not by the first reply.

Vo’s earlier review of the Muse app praised its permission design and activity feed, which fits the picture of a model built to act on your behalf more than to produce polished visuals. Meta has been courting agent builders since the Muse Spark API preview in July.

The design verdict is clear enough to keep Muse away from interface work. The agent ranking rests on one person’s task set and a partly AI-graded score, and her public write-up does not name the three winning tasks or give scores. It is a reason to run Muse on your own agent task and time it end to end, not a reason to switch.

End of article