Meta introduced the Muse Wonder series this week, expanding its agent ecosystem with two new frontier models: Muse Wonder and Muse Wonder Full. The architecture couples a 1-million-token context window with active memory and active focus, designed to maintain state across multi-hour tasks, orchestrate sub-agents, and recover from execution errors inside isolated environments. In public evaluations, Wonder Full captured top marks on developer benchmarks like DeepSWE and domain-specific tests including the Harvey Legal Agent Benchmark, handling end-to-end legal document synthesis, financial analysis, and large codebase refactoring.

Meta priced API access at $4.00 per million input tokens, $0.20 for cached input, and $20.00 per million output tokens, while restricting Wonder Full to Muse Code subscribers on its $50 Power and $200 Maximum monthly tiers. You can review technical documentation, pricing schedules, and benchmark evaluations directly in the official Meta announcement, the Meta AI Research deep dive, and the developer pricing portal.

While competitors remain focused on workplace seats and standalone developer APIs, Meta is pairing frontier agent reasoning with consumer-scale reach across its identity graph. For federal leaders and defense systems integrators, Wonder's arrival confirms that operational utility now demands reliable state persistence and sandboxed execution rather than clever conversational prompts alone.

What's Inside

Let's get into it.