Businesses weighing whether to let an AI assistant handle real work now have a new model to assess: Meta’s Muse Wonder and the larger Muse Wonder Full, released on October 12, 2026. Meta describes them as frontier models built for personal and professional agent tasks, from reading a lease to analyzing financial or health information and writing software. For an agency that looks after client websites and accounts, the release is a reason to review what such an agent would be allowed to touch before anyone connects one.
What Meta says Muse Wonder delivers
Meta calls Muse Wonder and Muse Wonder Full its most capable models to date for personal and professional agentic work, trained at its gigawatt-scale Prometheus data center. Both are available through the Meta Model API and Muse Code, with Muse Wonder Full limited to Muse Code subscribers, and Meta says the models will reach Muse app subscribers first before powering wider features. Meta’s developer blog lists Muse Wonder at $4 per million input tokens and $20 per million output tokens.
Those capability claims are Meta’s own, published alongside the release. Independent testing offers a second reference point. Artificial Analysis reported that "Muse Wonder Full (max) improves on Muse Spark 1.3 (max) across all ten Intelligence Index evaluations, gaining 8 points to 56", which placed the model among the higher-scoring systems it tracks. Its reported context window is 1 million tokens.
How does Muse Wonder handle domain expertise?
Meta says its evaluations include tasks in law, finance, and health, meant to test whether the models apply specialized knowledge to everyday problems. One example Meta cites is Harvey’s Legal Agent Benchmark, which requires a model to work with source documents and follow detailed instructions to produce legal work products. On the independently run version of that test, Artificial Analysis found that "Muse Wonder Full (Max) scores the highest on Harvey LAB-AA v1.1 with a score of 13.1%", followed by Muse Wonder (Max) at 10.0%.
What are Active Memory and Active Focus?
Meta groups the model’s behavior under two names. Active Memory is Meta’s term for carrying important context about a user across conversations and retrieving the details relevant to the current task, so earlier information resurfaces when it matters again. Active Focus describes how the model keeps working toward a difficult goal: trying alternative approaches, working through blockers, adapting to new information, asking for clarification when it needs it, and checking in before taking important actions. Meta says it measures focus through FocusBench and sustained execution through SWE-Marathon, which tests work on long-running software projects.
How strong is Muse Wonder at coding?
Coding stresses memory and focus at the same time, since it means tracking requirements, changing code across a codebase, and repeating cycles of testing and revision. Meta says it trained both models on diverse, long-running tasks to sustain progress through that process, and calls the result frontier coding performance for projects from small applications to complex software. In one demonstration Meta published, Muse Wonder builds a playable 3D kitchen game from customizable image inputs. Treat the demonstration as a vendor presentation rather than a measured result.
Do the skills carry over to unfamiliar tasks?
Meta tested how performance on tasks it targeted during training relates to performance on tasks it did not, comparing Muse Spark 1.3 and Muse Wonder Full as it scaled test-time compute. On the tasks targeted during post-training, Muse Wonder Full performed better, as expected. On unfamiliar tasks, both models improved at a similar rate, but Muse Wonder Full landed above the trend line its smaller predecessor would predict. Meta attributes this to a stronger base model, saying post-training gains carry over more effectively to unfamiliar tasks when the foundation is more capable. That matters for a personal agent, which has to handle combinations of goals, tools, and constraints nobody can fully anticipate in advance.
What safety work backs the release?
A personal agent takes actions on a user’s behalf and might handle context such as messages, calendars, and private documents, and that level of access demands a model that exercises discretion with private information and resists misuse. Meta says it trained the Muse Wonder family for robustness to adversarial instructions and careful handling of private information, and that the models carry its strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, while preserving access to frontier cyber capabilities for defensive work under its Superintelligence Scaling Framework. Meta describes Muse Wonder as its most aligned model to date. These are Meta’s statements, and Meta has said a full report with additional safety evaluations is still to be published, so the detailed evidence behind the alignment claim is not yet available to review.
What this means for your business
The practical change for a business is the kind of access a model like this asks for. An agent working across apps and a browser might handle context such as messages, calendars, and private documents to carry out multi-step tasks. Broader access is worth thinking through carefully. Before a client connects an agent to anything that holds customer data, the questions are about scope and control, not features.
One of those questions touches the website directly: whether an AI agent can actually read and use a site is a measurable thing. That is what an AI Agent Readiness check looks at, and SEOScanPro‘s AI Agent Readiness reports on it.
What to do now
If a client asks whether they should let a new personal AI agent into business accounts, this is general guidance an agency can walk through with them. Confirm the specifics against the vendor’s official documentation.
- Map the scope of data access. List exactly which inboxes, calendars, files, and systems the agent would be allowed to reach, and remove anything it does not need.
- Check whether the agent requires human approval before important actions, such as sending messages, making changes, or spending money. Meta says Muse Wonder is trained to check in before important actions; verify how that works on the surface you use.
- Ask how private documents are handled, including whether their contents are retained or used to train models.
- Review the vendor’s published safety documentation before relying on it. For Muse Wonder, Meta has said the full safety report is not yet out, so there is a gap to account for.
- Set identity and authorization rules up front. NIST’s National Cybersecurity Center of Excellence warns that "without strong identity, authorization, and governance, organizations risk data leaks, compliance failures, prompt injection, and unpredictable autonomous behavior".
FAQ
What is Muse Wonder and when was it released?
Muse Wonder is a frontier AI model from Meta, released on October 12, 2026, alongside a larger variant called Muse Wonder Full. Meta built them for personal and professional agent tasks, including long-running software projects, planning, and carrying context across sessions. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is limited to Muse Code subscribers.
Should a business let an AI agent access its calendars, inboxes or client documents?
That is a decision to review before connecting any agent, and it is general advice rather than a technical setup. NIST’s National Cybersecurity Center of Excellence warns that without strong identity, authorization, and governance, organizations risk data leaks, compliance failures, prompt injection, and unpredictable autonomous behavior. Check what data the agent can reach, whether it asks for approval before important actions, and how it handles private documents.
Has Muse Wonder’s performance been independently verified?
Meta’s benchmark figures are its own and were published with the release. Artificial Analysis, testing independently, placed Muse Wonder Full (Max) at 56 on its Intelligence Index and reported it scored highest on Harvey LAB-AA v1.1 with a score of 13.1%. Meta has said a full safety report is still to come.
Related coverage
- Meta Superintelligence Labs Ships Muse Spark: What It Means for Your Business - NewSunSEO
- Meta’s Muse Image and the Tagged-Photo Debate
- Meta Muse Spark 1.1 API Pricing and Developer Push
Sources
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
