On October 12, Meta released Muse Wonder and Muse Wonder Full, two new frontier models from its Superintelligence Labs aimed squarely at agentic work — tasks that unfold over hours, not seconds.
What's new: Active Memory and Active Focus
The two headline features distinguish these models from earlier entries in the Muse family.
Active Memory lets the model retain useful information across conversations and retrieve it when it becomes relevant again. This addresses one of the core frustrations with current AI assistants: the blank-slate problem where every new session starts from zero, forcing users to re-explain context they already gave yesterday.
Active Focus goes further: the model monitors its own progress through a difficult task, tries alternative approaches when it gets stuck, recovers from mistakes without abandoning the goal entirely, and — crucially — checks in with the user before taking consequential actions. That last point is a meaningful design choice. It's closer to how a diligent human collaborator works than how most AI systems behave today.
Availability and tiers
Muse Wonder is accessible now through the Meta Model API and Muse Code. Muse Wonder Full — the more capable tier — is initially limited to Muse Code subscribers. Forbes reports that the models support a one-million-token context window and can operate browsers, desktop applications, and mobile apps while managing long-horizon tasks.
How it benchmarks
Meta claims frontier-level performance across several evaluations. According to Yahoo Finance, on Artificial Analysis's Intelligence Index, Muse Wonder Full scored 56 — tying Claude Sonnet 5.5. The usual caveats apply: Meta's described use cases are demonstrations of intended capability rather than guarantees, and independent real-world evaluation at scale is still limited.
Where it fits
Muse Wonder follows Muse Spark, the first model in Meta's new family, launched earlier this year. The Muse branding ties into Meta's broader agent strategy — a push to position its AI infrastructure as the backbone for personal and professional agents across Meta's ecosystem and third-party integrations.
Unlike OpenAI and Anthropic, Meta has leaned toward open-weight and API-accessible models that give developers more control over deployment. For teams building agents that need persistent context and the ability to self-correct across long, multi-step workflows, Muse Wonder is the most capable option Meta has put forward to date.