Meta's Muse Wonder is strong as an AI agent but falls short on design, independent test shows

Meta released two models within the Muse family, named Muse Wonder and Muse Wonderful, positioned as agents for "personal and professional use" with frontier performance. Claire Vo, host of the How I AI podcast, decided not to trust the launch materials and ran both models, headless, inside Claude Code, against nearly 20 tasks from her How I AI benchmark. The result, published on October 14, 2026 in Lenny's Newsletter, sums up the episode in its own title: good for AI agents, bad for design.

For people who design products, this distinction matters more than any frontier performance claim in the launch materials.

How the test was set up

The How I AI benchmark covers nearly 20 tasks across five fronts: prototyping, agentic workflows, code, writing, and creative work. Vo ran the models headless via Claude Code and scored each task with an LLM-as-a-judge pass, complemented by her own voice notes. In other words: automated judgment plus human reading, not just an isolated benchmark number.

The Muse Wonder and Muse Wonderful duo was compared against models like Opus 5.5, Astra, Fable, and Sonnet, depending on the task category. The episode also covers Muse Code, the same family's coding offering (documentation at developer.meta.com/ai/lp/muse-code), but the contrast that gives the episode its name is a different one: agent versus design.

Where Muse Wonder really delivers

The model won in three of the benchmark's agentic tasks, categories that, in this piece's reading, matter because they simulate the real-world use of an assistant that needs to maintain context and make decisions over the course of a long task, not just respond to a single prompt. It's in this territory that Meta's "active focus mode," a feature announced to maintain attention on long tasks, proved real, albeit with caveats.

The "personal and professional agent" framing sold at launch holds up mainly here: workflow, email writing, and part of its performance as a voice assistant, where the model was placed side by side with Anthropic's Sonnet.

There's a contradiction that matters to anyone deciding on a tool for their team: Muse Wonder is fast at one-shot tasks and slow at agentic tasks, the opposite of what you'd expect from a model sold as a frontier agent.

Where it falls short: design is not a prompt commodity

This is where the angle that matters to people who design interfaces lives, not just to those who write agent code. According to Vo, Muse Wonder's design output "is not good," a direct assessment, with no middle ground, stated right in the episode's own summary. In vector illustration (SVG) and short video editing, the model was tested against competitors specialized in visual generation: Opus 5.5, Astra, and Fable.

The episode's overall verdict, good for AI agents, bad for design, indicates that it's precisely in the creative categories that Meta falls behind, despite a launch campaign that promised frontier performance across the board.

The risk Meta hasn't solved in design

Two days before Vo's test, the same newsletter published a conversation with Alex Cornell, Meta's design director responsible for Muse, about how the team designed "an AI agent people actually want to talk to": personality, voice, a sense of natural conversation. It's a genuine and valid design goal.

But it's a design goal for the agent's personality, not for the quality of the visual output that agent produces when it becomes a prototyping tool. This difference is the real risk for people who work with product.

A model can simulate judgment well (seem trustworthy, maintain context, "feel" like a colleague) and still have shallow aesthetic judgment when the task is actually to design something. A tool that simulates product decision-making well creates a dangerous illusion: the design team may confuse conversational fluency with delivery quality, and delegate visual decisions to a model approved only on the wrong dimension.

This is particularly dangerous for anyone in Brazil adopting a big tech's generative AI model assuming that "frontier," in marketing terms, covers every front equally. Vo's benchmark shows it doesn't: frontier in agent orchestration and frontier in visual judgment are different competencies, tested separately, with opposite results in the same model.

What this means for people who design products

Three questions are worth replicating before putting an agentic model on the design pipeline:

• Which model orchestrates my agent is a different decision from which model generates my interface. The test shows that, in Muse Wonder, the two competencies don't go hand in hand.
• Before using an agentic model to generate a visual prototype, it's worth running your own test, even an informal one, comparing the output against a tool specialized in design, not against the manufacturer's demo.
• A frontier performance claim without specifying the axis (agent, code, writing, creative) deserves suspicion: a model can lead in one benchmark and fall behind in the other three.

Vo closes the episode precisely by weighing where she will (or won't) use Muse Wonder in her own workflows, without publicly committing in this summary to which path she chose. It's the question worth bringing to your own product: before adopting the newest agent, what specific task did it solve better than what you already use, and what has it proven it still doesn't solve?