Meta launches Muse Wonder, focused on long tasks and computer use
Meta's new Muse Wonder model features Active Memory and Active Focus, with demos scheduling a calendar, planning a running route and generating a 3D kitchen game.
Meta just launched Muse Wonder and Muse Wonder Full, targeting coding and long-running agents. The pitch goes beyond writing code: • Active Memory retrieves relevant context across conversations. • Active Focus keeps the agent working through blockers, trying alternative approaches and adapting when the plan changes. • Coding and computer use bring those capabilities into actual workflows. In the video demo, Wonder checks a calendar and maps out a running route around personal preferences and scheduling constraints. Meta also shows Wonder building a playable 3D kitchen game from customizable image inputs. Wonder is available through Meta Model API and Muse Code. Wonder Full is available to Muse Code subscribers. Meta says the models compete with GPT-6 Astra and Claude Opus 5.5, based on its published benchmarks. Wonder is also planned for the Muse personal agent, starting with subscribers. The feature I’d test first is Active Focus. Can it recover from a failed step without losing track of the original task?