Meta’s Muse Wonder models are built for agents, and they’re its most capable yet

Meta has introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models to date. According to the official announcement, they are built for personal and professional agentic tasks, and Muse Wonder is available today.

What Meta announced

The Muse Wonder series builds on Muse Spark, the earlier model family from Meta’s rebuilt AI organization. Meta says it began overhauling its AI efforts 14 months ago and now describes the new models as state of the art for agents that handle everyday and expert work. The company points to tasks such as dealing with an insurance claim, reading a lease or making sense of personal finances. It also names law, finance and health as expert domains, and says the models offer frontier-level coding for projects ranging from games to business apps.

Two features define the release. Active Memory lets a model carry relevant details across conversations and bring them back when a task calls for them. Active Focus is meant to keep an agent working when a task gets difficult: it tries other approaches, recovers from errors, asks for clarification when needed and checks in before doing anything important.

Availability and pricing

Muse Wonder can be used now through the Meta Model API and Muse Code, Meta’s coding product. Meta says it will soon power more of the Muse assistant, with subscribers getting it first. The announcement does not give pricing. According to TechCrunch, which read Meta’s materials, API access to Muse Wonder costs $4 per million input tokens and $20 per million output tokens, and Muse Wonder Full is reserved for Muse Code subscribers and is not served through the API.

Meta’s research blog names benchmarks such as Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon, but the text of the blog gives no scores. TechCrunch reports that, in Meta’s own model card, both models score comparably to Anthropic‘s Claude Fable 5.1 and OpenAI’s GPT-6 Astra, and that Muse Wonder beats those rivals and Opus on Meta’s MemoryBench. These are the company’s own figures, and the full safety report is still pending.

Safety approach

Because a personal agent may work across messages, calendars and documents, Meta says it trained the models not to share private information without permission, to decline clearly harmful requests and to help strengthen cyber defenses. The company stresses that security teams and developers can still use the models to find flaws in their own code. Meta says it trains, evaluates and releases its most capable models under its Meta Superintelligence Scaling Framework, and that a full report with additional safety evaluations is coming soon. In its words, the release is “a major step toward personal superintelligence for everyone.”

Why it matters

Active Memory is the feature that turns Muse Wonder from a chat model into a personal agent, and it is also where the privacy questions concentrate. Meta says it trained the models not to share private information without permission. Reporting on how the Muse assistant handles context shows what that promise has to cover.

On October 3, WIRED reported that internal instructions extracted from Muse describe an hourly process that can build and update pages about people in a user’s life, such as family, friends and coworkers. Those pages can start sparse and grow over time. A Meta spokesperson told WIRED that Muse gathers context from public information and from material a person chooses to share, and that users choose which apps to connect, can disconnect them and can ask Muse to forget information. The instructions describe intended behavior, not a count of profiles built. The people described on those pages are not the ones who agreed to the product, and that is the harder question for privacy regulators.

Data use differs by route. According to a summary of Meta’s documentation and VentureBeat’s reporting, Muse Code subscriptions default to a Contributor mode, in which Meta can use prompts, code and outputs to train future models, and users can switch it off, though already submitted content is not covered. The same summary quotes Meta’s pricing page as saying Contributor mode is not available for Muse Wonder on the Model API, and that Standard mode content is not used for training. Developers on the API and subscribers in Muse Code therefore get different terms for the same model, and Wonder Full is available only on the second route.

Meta has said a fuller safety report is coming. Until it appears, the memory and training terms are what users can actually check.

What’s next

Meta says Muse Wonder will soon reach more of the Muse assistant, starting with subscribers, and that a full safety report is due. Independent testing of the benchmark claims will show how far the models match what Meta says.