Meta is making another push into the frontier AI race with Muse Wonder and Muse Wonder Full, two new models designed to handle complex tasks that require more than answering a question. Announced today, October 12, 2026, the models focus on coding, professional work, long-term memory, and the ability to keep working toward a goal even when a task gets complicated.

The launch marks a new stage in Meta’s effort to build AI systems that can act as personal and professional agents. Rather than simply responding to individual prompts, these models are designed to retain useful context across conversations, work through obstacles, and carry multistep tasks closer to completion.

Meta says Muse Wonder is available now through its Meta Model API and Muse Code platform, while the more advanced Muse Wonder Full is available to Muse Code subscribers. The company plans to bring Muse Wonder to its personal AI agent, Muse, starting with subscribers.

Muse Wonder focuses on getting work done

The two models build on the groundwork established by Muse Spark, Meta’s earlier model in its revamped AI lineup. Meta says the Wonder series delivers its strongest performance to date across evaluations covering agent capabilities, coding, and memory.

The intended applications go well beyond everyday chatbot conversations. Meta points to tasks such as navigating an insurance claim, interpreting a lease, making sense of financial information, and building a functioning application. The models are also designed to handle work in specialized areas such as law, finance, and health.

Coding is another major focus. Building an application often requires an AI system to understand a project’s requirements, modify multiple parts of a codebase, run tests, identify errors, and revise its work. Meta says Muse Wonder and Wonder Full are trained to sustain progress through these longer development cycles rather than treating each step as an isolated request.

That distinction matters as AI companies increasingly compete to build agents that can execute real tasks. A model might produce an impressive answer in a single exchange but struggle when a job involves several applications, changing requirements, and repeated attempts to solve a problem. Meta is positioning the Wonder series around those more demanding workflows.

Active Memory and Active Focus are the key features

Meta’s two concepts for making AI agents more useful over time

Meta is highlighting two capabilities behind the new models: Active Memory and Active Focus.

Active Memory allows the models to carry important information about a user across conversations and retrieve it when it becomes relevant. Instead of requiring someone to explain the same preferences, constraints, or background details repeatedly, an agent can use information from earlier interactions to inform its current work.

For a personal assistant, that could mean remembering a user’s preferences while planning a trip. In a professional setting, it could mean retaining project requirements and bringing them back into context when the user returns to a task days later.

Active Focus addresses a different problem: following through. Meta describes it as the ability to continue working toward a goal when difficulties arise, try alternative approaches, recover from errors, and adapt to new information. The models can also ask for clarification when necessary and check in before carrying out important actions.

Together, these features are intended to make an agent more useful across extended workflows. Memory helps it understand what matters; sustained focus helps it apply that context while working toward a result.

Of course, these are capabilities Meta describes and evaluates, not a guarantee that every task will be completed successfully. Real-world reliability will depend on the complexity of the job, the tools an agent can access, and how well it handles unexpected situations.

Meta says the models can compete at the AI frontier

Meta is also making a performance argument with the launch. Its research team says the Wonder models achieve state-of-the-art results across selected benchmarks for personal and professional agents, memory, and coding. The company’s evaluations compare them with other leading AI models, although benchmark results alone cannot establish which system will perform best across every real-world task.

One area Meta is emphasizing is the ability to generalize beyond familiar tasks. In its research discussion, the company argues that a model needs to handle unfamiliar combinations of goals, tools, and constraints—not simply excel at the types of problems emphasized during training.

That is particularly relevant for autonomous agents, whose work can take unpredictable turns. A coding agent, for example, might encounter an unexpected software error, while a personal assistant could discover that a booking or schedule conflicts with an earlier requirement. In both cases, the ability to adjust its approach can matter as much as its initial answer.

The models were trained at Meta’s Prometheus data center, which the company describes as a gigawatt-scale facility. Their release follows roughly 14 months of work to overhaul Meta’s AI efforts and develop a stronger foundation for more capable models.

Meta’s claims about competitive performance should still be viewed in context. The company has published benchmark results and says a more extensive safety evaluation report is forthcoming. Independent testing and experience with the models in everyday workflows will provide a clearer picture of how they compare with competing systems.

Availability, API pricing, and the Muse rollout

The two models have different access arrangements at launch.

• Muse Wonder: Available through the Meta Model API and Muse Code.
• Muse Wonder Full: Available to Muse Code subscribers, without API access at launch.
• Muse integration: Meta plans to introduce Muse Wonder into its personal AI agent, beginning with subscribers. The company has not announced a specific rollout date.

Developers using the Meta Model API can access Muse Wonder at a reported price of $4 per million input tokens and $20 per million output tokens. Those rates make the model a paid API offering rather than a freely accessible, open-weight release. Muse Wonder Full, meanwhile, is restricted to Muse Code subscribers under the initial availability plan.

Meta has not announced a public release date for Wonder Full through the API or confirmed a timetable for making the new models available to all Muse users. The company says it also intends to use Muse Wonder to power more intelligent experiences and develop specialized models for a broader audience.

That strategy puts the models at the center of Meta’s broader personal-agent ambitions. The company introduced Muse in September as an assistant designed to carry out tasks on a user’s behalf, working across connected apps and information sources rather than merely generating responses.

Safety becomes more important as AI agents gain autonomy

Giving an AI model the ability to work across messages, calendars, documents, and other tools creates a different set of risks from using a chatbot for general questions. An agent may encounter sensitive information or receive instructions embedded in material it reads, making privacy protections and resistance to malicious prompts important parts of its design.

Meta says the Wonder models were trained to avoid sharing private information without permission, refuse requests that facilitate harmful activity, and strengthen defenses against cyberattacks. At the same time, the company wants developers and security teams to retain the ability to use the models for defensive cybersecurity work, including finding vulnerabilities in their own code.

The company says it evaluates its most capable models under the Meta Superintelligence Scaling Framework and describes Wonder as its most aligned model to date. More detailed safety evaluations are expected to follow.

Those safeguards will be important as Meta expands Muse’s capabilities. An agent that can remember personal details and take action over an extended period needs to do more than reason effectively; it also needs to respect boundaries and handle sensitive information appropriately. Meta has outlined its approach, but the practical effectiveness of those protections will need to be assessed as the models see wider use.

Muse Wonder and Wonder Full therefore represent more than another model release. They show where Meta wants its AI strategy to go: toward systems that remember context, handle longer assignments, and act as capable assistants across personal and professional tasks. Whether the models can consistently deliver on that ambition will depend on how they perform outside the benchmarks—and how safely they behave when given real responsibilities.