Meta’s new ‘Muse Wonder’ model: good for AI agents, bad for design

Claire Voe introduces Muse Wonder as Meta's latest frontier intelligence model, available in two versions, Wonder and Wonderful. She explains her intention to evaluate Muse Wonder using her AI benchmark, focusing on applications in coding, knowledge work, and agentic workflows. She notes the model's branding centered on personal and professional agentic activities and mentions its comparison with models like GPT-6 Astra and Quad Opus 5.5, which she currently uses daily. She emphasizes that the review will assess Muse Wonder’s capabilities across a broad set of tasks from her benchmark, which includes nearly 20 different activities, judged via her voice notes and an LLN judge.

Claire discusses Meta's positioning of Muse Wonder, which claims state-of-the-art intelligence geared toward personal agents and domain-specific professional tasks, such as legal and financial work. She highlights Meta's emphasis on three 'frontier capabilities': domain expertise, active memory, and active focus. The model boasts superior memory for maintaining context over conversations and an 'active focus' mode that helps the model work on challenging goals by trying different approaches. Claire notes evidence from her testing that active focus can outperform on long-running tasks, supporting Meta’s claims.

Claire explains the process she used to test Muse Wonder, starting with subscribing to Muse Code, Meta's agent coding harness, which she admits found unimpressive. She downloaded it via an install script and used Claude Code to run it headlessly, avoiding any deep engagement with Muse Code itself. Her overall testing involved running the model through her comprehensive benchmark, covering tasks like coding, prototyping, agent workflows, creative tasks, and more. She describes her testing setup, including voice notes for scoring and the involvement of an LLN judge to evaluate the outputs, aiming for a first-look overview rather than a full benchmark.

Claire reports that Muse Wonder excelled in agentic work, outperforming on tasks like billing management, inbox triage, and long-term research, making it ideal for knowledge work and as a personal or professional voice agent. It performed well in coding tasks, placing it as a viable choice for backend development and architecture review, though Astra scored slightly lower. However, the model's creative and design outputs were disappointing; it struggled with prototypes, UI designs, and visual outputs, producing boilerplate, often poorly executed designs. Claire emphasizes that Muse Wonder’s speed is a major advantage for quick tasks but notes it slows down on complex, long-term agentic tasks.

Claire criticizes Muse Wonder’s creative outputs, including web design, SVG illustrations, and short-form videos. She finds its design sensibility outdated, with generic styles and poor execution, especially in high-visual-stakes projects like her Chat Purity website redesign, where it produced ugly, broken, or irrelevant images and layouts. She notes that for creative tasks requiring polish, other models like Opus 55 outperform Muse Wonder significantly. She also mentions that Muse Wonder’s SVGs and video edits are decent but not market-leading, suggesting that its strengths lie more in agentic and backend tasks than in creative design and art.

Claire summarizes her initial impressions: Muse Wonder is very fast and excels at short, agentic, and backend tasks, but falters considerably in creative design. While promising as an agentic voice and knowledge assistant, it’s not yet suitable for high-fidelity creative roles. She plans to continue testing its application in agent workflows and hopes that future iterations or skill enhancements might improve its creative capabilities. She concludes that Meta's model is progressing well and might become her preferred for specific in-loop tasks, but it is not yet ready to replace specialized creative tools. She encourages others to experiment with Muse Wonder in their workflows.