Oct 14, 2026 · 24 min
Meta’s Muse Wonder excels at agents, stumbles on design
Meta’s new ‘Muse Wonder’ model: good for AI agents, bad for design
The review tests whether Muse Wonder’s sustained effort and agentic strengths outweigh its unreliable frontend and creative work.
- 1Muse Wonder handled a complex customer-support workflow correctly, including account investigation, a duplicate refund, and a plan change.
- 2The model appears promising for sustained coding and knowledge-work tasks, though the coding sample was limited.
- 3Frontend redesigns, SVG work, and short-form video exposed major weaknesses in visual judgment and creative execution.
Don't miss
Muse Wonder successfully completes a complicated customer-support workflow involving account identification, a duplicate refund, a plan change, and prompt injection.
The brief
Claire Vaux puts Meta’s Muse Wonder and Wonderful through coding, knowledge work, customer support, frontend design, SVG creation, and short-form video tests.
The strongest result comes in customer support: Muse Wonder identifies an account issue, handles a duplicate refund and plan change, and resists prompt injection.
Muse Wonder looks better suited to sustained agentic work than quick creative judgment, with active memory and focused effort helping on longer workflows.
Its redesign of the ChatPRD website is visually poor despite brand guidance, while broken images, weak SVGs, and strange video audio undermine creative performance.
The verdict is a capable specialized agent model, not a universal favorite: fast on short tasks, slower on complex workflows, and still weak at design.