Quick Overview
In this benchmark review, How I AI host Claire Vo examines the release of Meta's frontier intelligence model family, Muse Wonder. The video responds to Meta's published performance claims against competing systems like Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1. Vo conducts hands-on testing across nineteen tasks to determine how the model handles coding, agentic autonomy, prototyping, and creative workflows.
Key Points
- 1.Meta's Muse Wonder Full ranked first out of eight models in agentic and knowledge work tasks on the How I AI benchmark.
- 2.The model ranked third out of eight in coding tasks, demonstrating capable performance in backend implementation and code auditing.
- 3.Muse Wonder fell behind rival models in front-end design and creative tasks, finishing sixth in prototyping and tied for seventh in creative evaluations.
- 4.Execution speed varied sharply, with one-shot writing and drafting running exceptionally fast, while multi-step agentic tasks required substantial run times.
- 5.Using Meta's coding harness required installing the Muse Code command-line interface, which can be run headlessly through external agents like Claude Code.
Summary
Claire Vo evaluates Meta's frontier model release, Muse Wonder, which is offered in two configurations: Wonder and Wonder Full. Meta positioned the model as a competitor to systems like Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra. Meta emphasized capabilities tailored toward agentic personal and professional knowledge work, supported by Active Memory for context retention and Active Focus for sustained task completion. To evaluate these claims, Vo runs Muse Wonder Full through the How I AI benchmark suite, covering nineteen tasks across agentic workflows, software development, document drafting, product prototypes, and creative assets.
Setting up the testing environment required obtaining a Muse subscription and installing Meta's proprietary agentic harness, Muse Code, via a terminal script available on Meta's developer site. Because Vo preferred desktop application interfaces and found the standalone terminal client unappealing, she connected Claude Code to run Muse Code headlessly. The benchmark evaluation gathered performance metrics scored through a blended system combining automated LLM judging with human review, incorporating recorded voice notes converted into structured rankings.
Muse Wonder Full performed best in agentic tasks, securing first place overall among eight tested models. In a challenging computer-use assessment involving an administrative SaaS console, the model investigated a problematic billing ticket. It successfully located the correct user account despite misleading linkage, refunded only one charge from duplicate entries, adjusted the subscription plan renewal, and ignored a planted prompt injection attempting to compromise permissions. It also achieved top marks for inbox triage, short concise email drafting matching Vo's preferred tone, voice assistant persona interactions, and backend code implementation.
The model performed adequately in coding, ranking third out of eight behind Sol 5.6 and Fable 5.1, proving dependable for backend architecture audits and pull request reviews. However, Muse Wonder Full struggled severely with front-end aesthetics and user experience design, placing sixth in interactive prototypes and tied for seventh in creative work. Wireframes and interface prototypes exhibited broken layouts, failed button interactions, and dated visual styling described as generic SaaS output with rounded corners and heavy serif typography. It failed to implement a requested dark mode theme for a developer incident tool and corrupted the layout of a web redesign test.
Creative tasks revealed additional weaknesses. While Muse Wonder Full successfully generated hand-coded SVG character illustrations across multiple emotional expressions, the stylistic quality fell behind competing frontier models. In a video editing test requiring the model to identify a clip and produce a captioned vertical short-form video, it selected an appropriate forty-second section and timed cut transitions well, but introduced corrupted audio, awkward visual framing, and unappealing caption placement. Operating speeds showed stark contrast: single-shot text drafting and prototype code generation ran rapidly, while complex agentic actions demanded long processing durations.
Meta Muse Wonder and Frontier Benchmark Claims
Meta introduced Muse Wonder in two sizes, Wonder and Wonder Full, positioning them as frontier models specialized in personal superintelligence and professional knowledge work. Meta's published benchmarks claim superiority over contemporary frontier models in agentic workflows, long-horizon software engineering benchmarks, and specialized domain expertise in legal and financial work. The release also highlights specific architectural capabilities dubbed Active Memory, which tracks context across conversations, and Active Focus, which supports sustained effort across complex, multi-step tasks.
Installation and Muse Code Experience
Accessing the model requires a paid Muse subscription, access to the Meta Model API, and the installation of Muse Code from Meta's developer portal. The installation process uses a terminal script to deploy an agentic coding harness. Because the terminal interface offered an unpolished user experience, the testing workflow was executed by having Claude Code run Muse Code headlessly, avoiding direct manual interaction with the command-line tool.
Agentic Strengths in Workflows and Backend Coding
Muse Wonder Full demonstrated top-tier performance across agentic evaluations on the How I AI benchmark, winning four out of five agentic tasks. In a complex computer-use task requiring navigation within an administrative application, the model resolved incorrect billing account attachments, addressed duplicate refund charges, and resisted an injected prompt instruction. It also excelled at triaging inboxes, drafting emails, managing voice assistant interactions, and completing backend software feature implementations.
Deficiencies in Design, Prototyping, and Media
The model underperformed significantly on design and visual tasks, producing dated user interface layouts with broken components, missing dark mode support, and generic typography. In front-end web design tests, including a website redesign evaluation, outputs suffered from flawed layouts and broken images. In creative assessments, Muse Wonder Full generated usable hand-coded SVGs but trailed competing models, and produced poor audio and visual layouts when cutting vertical short-form video clips.
The Bottom Line
The evaluation establishes that Meta's Muse Wonder Full is a highly capable frontier reasoning engine for backend development, workflow automation, and agentic administrative tasks. It nonetheless exhibits major flaws in visual aesthetic taste, front-end design execution, and multimedia processing when compared to Claude Opus 5.5. The review leaves open whether future updates, fine-tuning skills, or explicit prompt steering can fix the model's front-end styling deficiencies.
FAQ
What is Meta's Muse Wonder model and how does it perform in benchmarks?
Muse Wonder is Meta's frontier intelligence model family, available in Wonder and Wonder Full versions, designed for autonomous agentic workflows and professional knowledge work. In the How I AI benchmark suite, Muse Wonder Full placed first in agentic workflows and third in coding, but struggled in front-end design and creative tasks.
How did Meta's Muse Wonder model perform on complex computer-use agent tasks?
Muse Wonder Full ranked first among eight models on computer-use agent testing. In a billing support scenario, it successfully resolved duplicate charges, switched customer accounts, updated plans, and defended against a prompt injection attack.
What were the main weaknesses of Meta's Muse Wonder in prototype design evaluations?
The model produced unpolished, outdated visual interfaces characterized by broken clickable elements, lack of proper dark mode support, poor formatting, and unappealing typography.
How fast does Meta's Muse Wonder complete single-shot tasks versus multi-step agent tasks?
Muse Wonder operates very quickly on single-shot work, generating fifteen drafted emails in thirty-three seconds and completing prototypes in under four minutes. Conversely, multi-step agentic workflows require substantially longer run times to finish.
How was Meta's Muse Code agent harness configured and run during testing?
Muse Code was installed using an install script through the terminal from Meta's developer site, after which Claude Code was assigned to control and run Muse Code headlessly without manual terminal interaction.
Worth watching for
Software engineers, product managers, and technical leaders evaluating frontier large language models for autonomous agentic workflows, coding tools, and personal assistant implementations.
- meta
- muse-wonder
- llm-benchmarks
- ai-agents
- coding-assistants