Meta (NASDAQ: META) has rolled out Muse Wonder, a new frontier-class artificial intelligence model designed for its Muse personal agent application, as the company intensifies efforts to generate subscription income outside its traditional advertising business.
Released on October 12, 2026, the model—along with a more powerful variant called Muse Wonder Full—represents Meta’s most advanced offering yet for handling complex, multi-step real-world activities with greater adaptability than earlier systems.
Muse Wonder is intended to support demanding personal and professional agentic work, such as sorting through insurance claims, interpreting lease agreements, reviewing financial or health-related information, and assisting with software development.
Meta highlights two core capabilities: Active Memory, which allows the model to retain and retrieve relevant details across ongoing conversations, and Active Focus, which enables it to persist on difficult goals, explore alternative approaches when obstacles arise, recover from errors, and seek clarification or confirmation before taking significant actions.
Access for everyday users of the Muse app will initially be limited to those on paid plans.
Meta already offers a free tier of Muse with usage limits, alongside Power and Maximum subscriptions priced at $20 and $100 per month.
Muse Wonder is set to enhance the agent experience for these subscribers first, before any broader rollout.
Developers and builders can access the standard Muse Wonder model immediately through the Meta Model API (priced at $4 per million input tokens and $20 per million output tokens) and the Muse Code coding tool.
The larger Muse Wonder Full variant, which Meta positions as delivering higher performance on agentic and coding benchmarks, is restricted to higher-tier Muse Code subscribers on plans reported around $50 and $200 monthly.
Meta AI chief Alexandr Wang has described the development as aimed at making agents more reliable for everyday and specialized tasks, reducing instances where systems stall or require excessive user intervention.
The release follows a roughly 14-month overhaul of Meta’s AI research under the Superintelligence Labs, building on the earlier Muse Spark models that have powered the Muse agent since its U.S. launch in September 2026.
The agent itself runs tasks on dedicated virtual machines in Meta’s cloud and can connect to users’ apps and services to perform actions like form-filling or bookings, while operating independently of ongoing chat sessions.
On safety, Meta states that Muse Wonder provides its strongest protections to date against chemical, biological, radiological, nuclear, and cyber misuse, while still permitting defensive security research.
Training emphasizes careful handling of private information, resistance to harmful requests, and alignment with the company’s scaling framework.
This emphasis comes after reports of user complaints that the Muse agent accessed personal messages or data without clear authorization; Meta has disputed certain claims and indicated it is working to improve user education around permissions and prior approvals.
The move underscores Meta’s strategy to convert heavy AI infrastructure investments into recurring revenue through subscriptions and potential future commerce features, amid competition with models from OpenAI and Anthropic. Independent evaluations place the Wonder series near the top of aggregated capability rankings, though Meta notes a fuller safety report is forthcoming.