Meta announced two new models on October 12: Muse Wonder and Muse Wonder Full. Muse Wonder is available through the API and through Muse Code. Muse Wonder Full is restricted to Muse Code subscribers only. Both models carry two named features, Active Memory and Active Focus, built for agentic tasks, according to Meta's announcement at about.fb.com.
The split matters for anyone who built integration plans assuming a single access path. A model gated behind a subscription product behaves differently from an open API endpoint in terms of rate limits, auth flow, and failure modes. Teams that provisioned API keys expecting Muse Wonder Full to show up there will not find it. That is a structural fact, not a bug, and it changes which credentials and which quota pools matter for any pipeline touching these models.
Meta's announcement confirms the names, the two-tier access structure, and the two feature labels. It does not disclose pricing, rate limits, latency figures, context window size, or rollout regions. It does not explain what Active Memory stores, for how long, or where. It does not say whether Active Focus changes token consumption or billing. Anyone building monitoring or alerting around these models is working from a product name and a feature label, not a technical spec. That gap is the thing to flag before writing any dashboard query against API responses.
On-call engineers should check which authentication path their service actually uses before assuming parity between Muse Wonder and Muse Wonder Full. If a service calls the API directly, it can reach Muse Wonder but not the Full variant, full stop. If it routes through Muse Code, subscription status becomes a dependency worth tracking in the same way a license expiration is tracked, because a lapsed subscription could silently downgrade which model answers a request. Active Memory, if it persists state across calls, introduces a new failure class: stale or leaked context between sessions, which standard stateless retry logic will not catch.
Platform operators should also verify whether existing error-handling and timeout budgets were tuned for a different model family, since agentic features like Active Focus imply longer-running, multi-step calls rather than single-shot completions. A timeout set for a quick inference call may cut off a legitimate multi-step agentic task midway. Nothing in Meta's announcement quantifies expected latency for either model, so teams have no baseline to compare against until they run their own load tests. That absence of published numbers is itself the operational risk worth logging now, before traffic shifts onto either model.
The near-term decision is narrow but concrete: confirm which of the two models any given integration can legally and technically reach, and confirm it against the subscription status, not against assumptions carried over from prior Meta model releases. Muse Wonder Full is not available to plain API callers as of October 12. That single access rule, more than the feature names, is the fact to encode into any routing logic before the next on-call shift starts.