Meta Launches Muse Wonder and Muse Wonder Full, Its Most Capable AI Models Yet
Meta released Muse Wonder and Muse Wonder Full on Monday, October 12, calling them its most capable models to date and its first entry into the frontier tier of AI systems since Alexandr Wang restructured its AI efforts. The models are designed for personal and professional agentic tasks, from navigating insurance claims and understanding leases to building software.
 |
| Credit: Meta |
Muse Wonder is available through the Meta Model API and Muse Code, priced at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is available only to Muse Code subscribers and is not served through the API.
Neither model is being released with publicly downloadable weights, a departure from Meta's earlier Llama strategy of letting organizations run models on their own infrastructure.
The models were trained at Meta's Prometheus data center in Ohio and support context windows of approximately one million tokens, accepting text, images, audio, and video as inputs.
Independent benchmarking organization Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index, tying Anthropic's Claude Sonnet 5.5 and surpassing OpenAI's GPT-6 Astra and Google's Gemini 4 Argon, which each scored 53. The standard Muse Wonder scored 52.
What Meta emphasizes most is not conventional chatbot performance.
The company built the Wonder models around two capabilities it calls Active Memory and Active Focus.
- Active Memory allows an agent to carry context across conversations, retaining details about a user's preferences, history, and ongoing projects.
- Active Focus keeps an agent working toward a goal when a task encounters obstacles, allowing it to try alternative approaches, recover from errors, and ask for clarification before taking consequential actions.
Meta says these features are what make the models suitable for real work that unfolds over days or weeks.
Meta's research blog states that the models were trained to avoid sharing private information without permission and to refuse requests that assist harmful acts.
"Muse Wonder is our most aligned model to date," the company wrote, adding that a full safety report would be available soon.
The privacy record behind the assurances
Meta's privacy history makes those assurances difficult to accept at face value. The company is operating under a $5 billion FTC penalty tied to the Cambridge Analytica scandal and a 2020 consent order that required expanded privacy protections and independent third-party assessments.
In August 2026, Meta agreed to pay up to $17.1 billion to settle claims from 47 states that it designed Facebook and Instagram to be addictive to young users. As part of that settlement, the company agreed to some platform changes for minors. It did not agree to stop collecting data from young users or to stop using that data for targeted advertising.
The Muse agent, which the Wonder models will power, has already produced a series of privacy incidents since its September launch.
A Toronto tech reviewer named Matt Robb had his home address shared with a stranger on Facebook Marketplace after granting Muse permission to handle Marketplace inquiries on his behalf.
CNET reported that Muse requests extensive access to personal data, including passwords, full disk access, email messages, calendar, contacts, Notes, and WhatsApp, and that by default Meta uses interactions with Muse to train its AI models unless users manually opt out.
Researchers at the University of St. Gallen in Switzerland found that Muse maintains detailed profiles not only of users but of every person with meaningful ties to them, including people who are not Muse users and have no say in being profiled.
The profiles include birthdates, addresses, occupations, and important dates.
A Wired report found that Muse updates these files once an hour with an internal instruction to maintain "a page for every person in the user's life."
Meta updated its privacy policy in late 2025 to clarify that prompts and interactions with its AI services can be used to personalize ads across Facebook and Instagram.
A coalition of 36 privacy and consumer protection groups asked the Federal Trade Commission to investigate, arguing the policy represented an expansion of surveillance-driven marketing. The FTC has not acted.
What users are actually trusting
The concern about Meta's privacy commitments is not speculation about a distant scenario. It is the logical endpoint of a business model that has already normalized the use of personal data for advertising across every product Meta operates. The company's revenue depends on advertising, and advertising depends on data. Meta has a demonstrated record of expanding data collection when it serves that business, and its privacy policies are not contracts. They are documents the company can revise.
What Meta is asking users to accept with Muse Wonder is different from what it asked of them with a chatbot.
An agent with Active Memory that navigates insurance claims, reads leases, and builds software is an agent with sustained access to the most sensitive categories of personal information.
The privacy safeguards Meta describes are training objectives and policy commitments, not architectural guarantees. A model trained to avoid sharing private information can still be reconfigured.
A policy that promises not to use agent data for advertising can be changed with a notification.
Meta's own privacy policy already contains the mechanism. The company told users in 2025 that AI interactions would be used to personalize ads. That decision required no affirmative user consent beyond continued use of the platform.
The same logic applies to an agent that holds months of context about a user's health, finances, and relationships. The data is already on Meta's servers. The question is what the company does with it next.
None of this means Muse Wonder is unsafe to use. The benchmarks are strong, the agentic capabilities are real, and Meta has published more detail about its safety training than it did for earlier models.
But readers evaluating whether to hand an AI agent access to their email, calendar, and documents should understand what they are actually trusting. They are trusting a company that paid $5 billion for privacy violations, agreed to a $17.1 billion youth safety settlement without giving up its data collection practices, and launched an agent that shared a user's home address with a stranger within weeks of release.
Meta says Muse Wonder will soon power experiences in Muse and roll out first to subscribers. The models are available now. The privacy questions they raise will not be settled by a blog post.
