Meta has released the Muse Wonder models, a two-model AI lineup aimed at the company’s Muse agent and coding tools. The useful part for prospective users is not Meta’s “frontier” rhetoric. It is the access split: Muse Wonder is available through paid Muse subscriptions, Meta’s Model API and Muse Code, while Muse Wonder Full is limited to paid Muse Code plans.
Meta announced Muse Wonder and Muse Wonder Full on October 12. The company describes both as its most capable models for personal and professional agent tasks, a claim based on evaluations it published itself. The available material contains no independent benchmark results, technical specifications or model-card details sufficient to verify those comparisons.
How can developers access Meta Muse Wonder and Muse Wonder Full?
According to Meta’s announcement, Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.
CNBC reports that access to Wonder through Meta’s Muse personal-agent app requires a monthly subscription priced at $20 or $100. It also reports that Muse Code plans cost $50 or $200 a month, with Wonder Full confined to that software-engineering-focused service. Meta had not listed API pricing in the material cited, though AI chief Alexandr Wang told CNBC it would cost more than the standard subscriptions.
The distinction is meaningful. Meta is not offering equal access to both models: the fuller variant is reserved for people paying for Muse Code, while Wonder has the broader set of currently reported routes into Meta’s services.
What does Meta say the new models do?
Meta says Wonder and Wonder Full target tasks for personal and professional agents, including work that needs contextual recall, sustained task execution and coding. Its terms for two of those claimed capabilities are “Active Memory” and “Active Focus.”
In Meta’s description, Active Memory means holding onto useful context across conversations and retrieving it when a later task calls for it. Active Focus refers to pursuing a complex task by trying other approaches, handling blockers, adapting to new information and requesting clarification where needed. Wang told CNBC the work was intended to help agents keep moving through setbacks instead of stopping when a task runs into trouble.
Meta also says it has its strongest safeguards yet against cyber misuse in the paid model. That is a company assertion, not an independently established safety result.
When will Muse Wonder power Meta’s Muse agent?
Meta says Wonder will soon begin powering Muse experiences, with subscribers first in line. It also says Wonder will support artifact generation and other processes that require more intelligence. The company has not supplied a specific rollout date or detailed product schedule.
Meta says internal evaluations place the models alongside leading systems from OpenAI and Anthropic. The Verge reported that the comparison with Google’s Gemini 4 Argon remained unclear. Until outside testing is available, the performance claims should be read as launch-day claims from the vendor, not a settled league table.
This story draws on original reporting from The Verge.
