Meta has launched its new artificial intelligence models, Muse Wonder and Muse Wonder Full, as part of an effort to develop AI agents capable of handling complex personal and professional tasks over longer periods rather than simply answering questions or generating content.
The company announced the launch of the two models on October 12, 2026, with Muse Wonder made available to developers through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
What’s New in Muse Wonder?
Meta is focusing on making its new generation of AI models more capable of handling tasks that require multiple steps and longer periods of work.
According to the company, Muse Wonder is designed for practical applications such as understanding contracts, analyzing financial information, assisting with insurance claims, software development, and other tasks that require the use of tools and sequential decision-making.
The family includes two models: Muse Wonder and Muse Wonder Full, both of which Meta describes as among its most advanced models to date.
Active Memory for Maintaining User Context
One of Muse Wonder’s key capabilities is what Meta calls Active Memory.
The technology allows the model to retain important information about a user across multiple conversations and retrieve relevant details when needed. The goal is to reduce the need for users to repeatedly provide the same information in every conversation.
This concept goes beyond simply retaining a conversation history. Meta says the model attempts to identify information that is relevant to the task and retrieve the appropriate details while working.
Active Focus Helps the Model Continue When Problems Arise
Meta is also introducing Active Focus, a capability designed to allow the AI agent to continue working when it encounters an obstacle instead of stopping immediately.
According to the company, the model can try alternative approaches, deal with errors, adapt to new information, ask for clarification when necessary, and verify important actions with the user before carrying them out.
This capability could be particularly important for tasks involving a long sequence of steps, such as working with multiple documents or completing complex software-development tasks.
Muse Wonder Is Also Designed for Developers
Meta is making Muse Wonder available to developers and builders through the Meta Model API and Muse Code, while Muse Wonder Full is available through Muse Code subscriptions, according to the company’s announcement.
The cost of using Muse Wonder through the Meta Model API is $4 per million input tokens and $20 per million output tokens, according to Meta’s developer documentation.
The capabilities are primarily aimed at applications that rely on AI agents capable of completing multiple steps rather than simply generating text-based responses.
Meta Takes on OpenAI and Anthropic
The launch of Muse Wonder is part of Meta’s efforts to strengthen its position in the advanced AI model market, amid competition from companies such as OpenAI and Anthropic.
Meta said the two models deliver advanced performance on personal and professional tasks, while reports have indicated that the company internally compares their performance with competing models.
Such comparisons should be viewed in the context of specific benchmarks and testing conditions rather than as a universal judgment that one model is superior to another in every use case.
Independent evaluations indicate that Muse Wonder and Muse Wonder Full have made notable progress compared with some earlier Meta models, although results vary depending on the benchmark and level of reasoning involved.
Muse Wonder Comes to the Personal Assistant Experience
Meta’s plans are not limited to developers. The company announced that Muse Wonder capabilities will also be coming to the Muse user experience, with availability beginning for subscribers.
This move could turn the model into part of a personal assistant experience capable of handling longer and more complex tasks rather than relying solely on traditional conversational interactions.
Privacy and Security Face New Challenges
As AI agents become more capable of handling messages, documents, calendars, and actions on behalf of users, data protection and permission management are becoming increasingly important challenges.
Meta says it has trained the Muse Wonder models to avoid sharing private information without permission, reject certain harmful requests, and strengthen their cybersecurity capabilities. The company also said it plans to provide a more detailed report on its safety evaluations.
This highlights one of the major turning points for AI agent technology: as systems move from providing information to actually taking actions, the accuracy of the model, the limits placed on its permissions, and the protection of user data become just as important as its ability to reason and complete tasks.
A Step Toward More Autonomous AI Agents
Muse Wonder reflects a broader shift in the AI industry toward models that can do more than simply answer questions. These systems are increasingly designed to retain context, plan tasks, use tools, handle obstacles, and continue working until they reach a result.
For Meta, the launch of the two models represents an attempt to move into a more advanced phase of personal and professional AI assistants while connecting these capabilities to the company’s broader AI ecosystem.
Ultimately, the ability of these models to deliver reliable and safe results in everyday use will be a key factor in determining the success of this approach.