Meta introduced Muse Wonder and Muse Wonder Full on October 12, targeting AI that can carry complex work through multiple steps—not just answer a prompt.
Muse Wonder is available in preview through the Meta Model API and in Muse Code. Full is currently restricted to Muse Code’s Power Usage and Maximum Usage subscriptions; it is not a generally available API model.
Memory and focus for sustained work
The release centers on two capabilities Meta calls Active Memory and Active Focus. Active Memory describes retrieving relevant context from earlier conversations. Active Focus describes persisting through difficult tasks by trying alternatives, working through blockers, and recovering from errors.
Meta positions both models for sustained software development and personal or professional workflows. These are capability claims, not guarantees of reliable autonomous execution. Developers should not assume that every API integration automatically includes cross-session memory; the surrounding application still determines which context, tools, and stored information are available.
Availability and API pricing
For developers, Muse Wonder offers a one-million-token context window. Meta lists API pricing at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens.
The consumer Muse rollout is still upcoming, starting with subscribers. That separates today’s developer and Muse Code access from the planned broader consumer experience.
Why it matters
For app builders such as SunMarc, this adds another model to evaluate for repository-wide changes, debugging loops, and document-heavy workflows. The useful test is whether it completes real tasks with fewer retries at an acceptable cost—not simply where it lands on a leaderboard.
The release extends the sustained-work direction we covered in Meta’s Muse Spark 1.3. Keeping requirements intact through edits, tool results, and repeated tests matters more to a shipped app than a strong first response.
Meta’s model card notes that performance depends on prompts, tools, serving configuration, and the agent harness. Teams should compare completed tasks, retry counts, and total cost in their own workflow before treating launch demonstrations as production evidence.