Meta Launches Muse Wonder: Availability, Active Memory and Agent Reliability
Meta launched Muse Wonder on October 12. Explore API access, Muse Code plans, Active Memory and the reliability checks that matter for developers.
Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, expanding its model lineup for coding and agent tasks. The release puts a practical question in front of software teams: can an agent preserve the right context and complete a longer workflow without creating new operational problems?
Illustrative computing workspace; this photograph does not show Muse Wonder or Muse Code. Photo: Howard Bouchevereau / Unsplash, Unsplash License.
What developers can access now
Meta's announcement makes the launch surfaces clear: Muse Wonder is available through Meta Model API and Muse Code; Wonder Full is offered to Muse Code subscribers. A future rollout into the consumer Muse experience is a separate promise, beginning with subscribers.
The developer launch guide narrows Full access to the Power Usage and Maximum Usage plans. It lists Wonder API pricing at $4 per million input tokens and $20 per million output tokens. The model page also lists cached input at $0.20 per million tokens and a one-million-token context window. These are published launch terms checked on October 13, rather than a prediction of a team's total bill.
TechCrunch covered the release on October 12 at 10:56 a.m. PDT. The Verge's same-day report distinguishes the developer launch from the planned Muse rollout.
Active Memory and Active Focus address different problems
In its research explanation, Meta describes Active Memory as carrying important context between conversations and retrieving relevant details. Active Focus concerns persistence: trying other approaches, adapting to obstacles and checking with the user before important actions. The company evaluates sustained work through tasks including long-running software projects.
The distinction matters. Remembering a previous requirement helps an agent choose what to do; continuing through a failed build helps it finish. Neither capability alone establishes that the final change meets the requirement. A development workflow still needs evidence that the code builds, the intended behavior works and the deliverable actually exists.
How to evaluate the release in a real repository
Makezaa analysis: a useful trial should start with a bounded change and a known acceptance condition. For example, ask the agent to add pagination to a test application, preserve a stated accessibility requirement and recover from an intentionally failing check. Compare the final behavior with the requirement, rather than counting how many tool calls it makes.
Test memory separately. Introduce a requirement, correct it later and check whether the agent uses the correction. Repeat with unrelated information that should not influence the task. This exposes the difference between recalling more context and selecting the correct context. A remembered preference also should not become permission to deploy, spend money or disclose a document.
Cost measurement should include unsuccessful attempts and retries. A lower token price can be outweighed by repeated work, while a more expensive call may finish a task with fewer interventions. Record completed tasks, reviewer time and actual billed usage together; a single headline benchmark cannot answer that operational question.
What remains to be established
Meta presents strong benchmark results, but this article does not independently validate its rankings or treat them as a guarantee for every codebase. Its research post says a fuller safety evaluation report will follow. Capability claims and future reports should therefore remain distinct from observed production behavior.
Before a broader rollout, teams should define where memory is stored, who can inspect or correct it, and how tool permissions are enforced. An agent that remembers a project well still needs a reliable boundary around external actions. Our practical guide to AI agents for small teams explains how verified results and explicit permissions fit into that workflow.
The immediate developer value is a new model option with stated access paths and a clear emphasis on longer tasks. Whether it improves a particular team's delivery process is best established through repeatable trials with visible outcomes.