Meta launches Muse Wonder for long-running AI agents
Meta’s first frontier-class models target persistent memory, computer use and complex coding, putting its consumer-agent strategy on a new footing.
Meta has released Muse Wonder and Muse Wonder Full, its first models explicitly positioned as frontier competitors to leading systems from OpenAI, Anthropic and Google. The launch is the clearest result yet of Meta’s 14-month restructuring of its artificial-intelligence effort under Meta Superintelligence Labs.
Built for work that continues
Meta says Muse Wonder is available through its Model API and Muse Code, while Muse Wonder Full is restricted to Muse Code subscribers. Both models support a one-million-token context window and are designed for long-running, multi-step tasks. The company highlights “Active Memory,” which carries relevant information across conversations, and “Active Focus,” which lets a model change strategy, recover from errors and ask for clarification before taking consequential action.
The target workloads go beyond chat: navigating insurance claims, understanding leases, handling financial information, building applications and operating browsers or desktop software. Meta’s model page also describes parallel delegation to sub-agents and continued execution over hours.
A strategic shift for Meta
The release matters because Meta is no longer presenting itself mainly as the distributor of other laboratories’ advances. Muse Wonder gives the company a proprietary top-tier model to power Muse, its personal-agent product, while the API gives developers an independent route into the stack. TechCrunch reported pricing of $4 per million input tokens and $20 per million output tokens for Wonder, placing it near other premium models.
Meta has not announced open weights for the Wonder family. That leaves a sharp distinction between its frontier research strategy and its earlier open-model identity. The company also says that a fuller safety report will follow, an important gap for systems designed to access messages, calendars, documents and other private data.
The significance is therefore broader than another benchmark launch: Meta is trying to connect frontier intelligence, persistent personal context and billions of consumer touchpoints in one product strategy. Whether users will trust an agent that remembers enough to be useful remains unresolved.