What changed
On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full. Meta says Muse Wonder is available now through its Model API and Muse Code, whereas Wonder Full is available to Muse Code subscribers and is not announced as a general Model API offering. Meta plans to bring Wonder to its consumer Muse agent, starting with subscribers, but that rollout is future rather than completed. Both models introduce what Meta calls Active Memory, retrieval of relevant user context across conversations, and Active Focus, sustained multi-step work with retries and clarification. The vendor claims frontier results in coding, long-running agent tasks and professional-domain evaluations.
Why it matters
Developers evaluating agent models must distinguish an API-accessible model from a premium model restricted to an interactive coding product. The promise of persistent memory and longer-running task execution is consequential for coding agents and SaaS products, but it also increases the importance of permission boundaries, personal-data retention, auditability and error recovery. Teams cannot assume Meta's strongest reported benchmark numbers are available via the same API, or that self-reported benchmark wins transfer to their own tasks.
Two models, different access routes
Meta lists Muse Wonder for developers through Meta Model API and Muse Code. Wonder Full is offered to Muse Code subscribers; Meta has not confirmed general API access for Full. Builders should verify model identifiers, current rate cards, context limits and feature availability in their own account before planning a migration. No verified public Wonder-specific API price was established in this research.
Active Memory and Active Focus are agent behaviours, not proof of reliability
Meta describes Active Memory as carrying context across conversations and recalling relevant details, and Active Focus as trying alternate approaches and recovering from errors on longer tasks. Its research blog points to FocusBench and SWE-Marathon, along with law, finance and health evaluations. These are vendor claims; independent long-horizon success rates, memory privacy behaviour and cost-per-completed-task remain to be established.
Consumer Muse integration is promised, not complete
Meta says Wonder will later power the Muse assistant and roll out to subscribers first. This is distinct from today's Model API and Muse Code access, so existing Muse users should not assume they are already using Wonder.
Safety and data-handling claims need separate validation
Meta says it trained the models to protect private information and resist misuse, and references its Superintelligence Scaling Framework. Its full additional safety evaluation report was still forthcoming at announcement. Agent deployments involving personal documents, connected accounts or purchases require their own permission scopes, human approvals and monitoring.
What independent reporting verifies
Reuters reported the October 12 model launch and the competitive positioning against other frontier labs. The availability distinction is documented directly by Meta; benchmark leadership, safety performance and productivity improvements are not independently demonstrated by the launch reporting.