The News, Explained
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026 for personal and professional agentic tasks. An agent in this context does more than answer a single prompt: it can use documents and tools to carry a multi-step job forward. Meta cites navigating an insurance claim, understanding a lease, making sense of financial information and building an app as examples. Those use cases and performance descriptions are Meta’s claims, not independent evaluation results. Source
Access differs between the two models. Muse Wonder is available through the developer-facing Meta Model API and the Muse Code coding tool. The larger Muse Wonder Full is available to Muse Code subscribers. Meta also plans to use Muse Wonder in its Muse AI service, with subscribers receiving it first. The announcement does not state API pricing, rate limits, supported regions, context limits or the Muse Code subscription price. Source
The models use concepts Meta calls Active Memory and Active Focus. Active Memory carries important information across conversations and retrieves details that matter to the current task. It could, for example, apply requirements collected over several discussions to later work. Active Focus refers to continuing through difficult steps, trying other approaches and recovering from errors. Meta says the models ask for clarification when information is missing and check with the user before important actions. Source
Memory and tool use can add convenience while increasing privacy risk. Because an agent may handle messages, calendars and documents, Meta says it trained the models not to share private information without permission and to reject requests that would assist actually harmful acts. It also says defensive security work, such as finding flaws in a developer’s own code, remains supported. These are company-stated training goals; the announcement does not provide failure rates or results from attempts to bypass the safeguards in deployed use. Source
Meta says the models were evaluated and released under its Superintelligence Scaling Framework. A full report with additional safety evaluations is still scheduled for later publication. What is available now is the company’s stated safety approach and the access routes, rather than a complete set of measured risks and operating limits. Source
OYOPICK’s Take
The practical change is that a model designed for sustained agent work is moving beyond a research description into an API and coding tool that are available now. Carrying context across conversations and trying alternatives after errors could reduce the need for people to repeat instructions or restart interrupted work. Its value should be measured against earlier models on completion rate, correction count and the amount of human intervention required for the same task.
Memory cannot be judged only by how much a system remembers. Users also need to know what is stored, how long it is retained, where it can be reused and whether each item can be reviewed and deleted. We hope Meta puts memory scope, deletion controls and approval steps before tool actions in one clear place across Muse and the API.
Meta’s emphasis on legal, financial and health tasks deserves caution while the full safety report remains unpublished. A useful assistant in high-impact work needs evidence beyond average accuracy: it should report how often the model detects its own mistakes, what privacy-leakage tests show and when it stops before an important action. Those results, together with pricing and usage limits, will determine whether developers and users can make a grounded comparison of Muse Wonder.