Meta says Muse Wonder has caught up to the AI frontier, but user results remain unproven

Artificial Analysis, an independent evaluator, scored Muse Wonder 52 and Muse Wonder Full 56 on its Intelligence Index v4.3 at maximum reasoning effort. Its comparison also listed Muse Spark 1.3 at 48. The published scores provide evidence beyond Meta’s own account, but they measure performance on an index—not how often the models give correct, complete and understandable help with a lease, insurance claim or legal question.

The scores are useful as a comparison within that evaluation. They cannot, by themselves, tell a person whether to rely on a model’s reading of a complicated document, or what to do when important information is missing. Meta’s broad claim about reaching the frontier asks more of the evidence than an index score can establish.

Who can use the models now?

Availability differs between the two models. Meta says Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers. Meta also says Wonder will power experiences in its Muse agent later, starting with subscribers. That integration is planned, not available now.

That distinction matters to people hoping to use an agent for an insurance claim or lease. Access to a model through a developer-facing service or coding product does not mean the same assistance is already available through Meta’s agent. Engadget reports that Muse Code tiers costing $50 and $200 per month provide access to Wonder Full. Those prices describe one route to the model, not whether it will be affordable or useful for a particular person.

Meta’s examples point to tasks where a wrong answer could have consequences. But the company’s announcement does not document results from people who have used the models to navigate an insurance claim or understand a lease. Before treating access as practical help, users need evidence about accuracy, costs and the experience of correcting errors.

What would make the models meaningful for workers and users?

Meta says the models add Active Memory, which carries relevant information across conversations, and Active Focus, which helps an agent continue working toward a goal, recover from errors and ask for clarification. Those features could matter when a task takes several exchanges. They also make questions about control important: what information persists, how users manage it, and what happens when a consequential detail is wrong?

The company says the models were designed to handle private information carefully and resist misuse. Its announcement says a fuller report with additional safety evaluations will be available later. Until that report and other evidence arrive, these safeguards remain Meta’s description of the models’ design, not independently established protection in everyday use.

The professional-work pitch also needs evidence about people’s experience. The available sources establish no measured effects on workers’ jobs, pay, workload or time, and provide no accounts from people using the models for the tasks Meta names. A model’s stated focus on professional work does not show whether it saves time, shifts responsibility or creates extra work to check its output.

Meta should publish independent, task-specific evaluations for legal, financial and health-related work, including how often the models make consequential errors, alongside the promised safety report and clear information about memory controls. Meta controls the rollout into its products; it should make that evidence available before asking people to rely on the models for decisions affecting their money, rights or work.