Muse Wonder Explained: Meta's Memory-First Frontier AI

Muse Wonder is Meta's new top-tier model family, and its defining idea is memory. Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models for personal and professional agentic tasks. The larger design sits above the low-latency Muse Spark line that Meta has shipped through the year.

• Released: October 12, 2026, as Muse Wonder and Muse Wonder Full.
• API price: $4 per million input tokens and $20 per million output tokens for Muse Wonder, as reported by TechCrunch.
• Availability: Muse Wonder on the Meta Model API and in Muse Code; Muse Wonder Full for Muse Code subscribers.
• New features: Active Memory and Active Focus.

What Is Muse Wonder's Active Memory?

Meta's announcement describes two named capabilities. Active Memory keeps important context about a user across conversations and retrieves the relevant details for the task at hand. Active Focus is about persistence: the model keeps working toward a goal by trying alternative approaches, working through blockers, adapting to new information, asking for clarification and checking in before important actions.

Taken together, these address a real limitation of current assistants. A model that forgets everything between sessions has to be re-briefed each time, and a model that abandons a task at the first obstacle needs constant supervision. Memory and persistence are the two properties that make an agent useful over days and not only minutes.

Meta names three benchmarks in its post: Harvey's Legal Agent Benchmark for legal work products, FocusBench for sustained attention, and SWE-Marathon for long-running software projects. The text of the post does not publish scores, so I will not quote any.

How Does Muse Wonder Compare on Price and Benchmarks?

TechCrunch reports that Muse Wonder costs $4 per million input tokens and $20 per million output tokens on the Meta Model API, and that Muse Wonder Full is offered only to Muse Code subscribers and is not served through the API.

On capability, the claims need careful attribution. According to TechCrunch, Meta's model card says the Wonder models score comparably to other current frontier models, and higher on MemoryBench, a benchmark built around continual learning and memory construction. Those are Meta's own results.

There is one early independent data point. VentureBeat's October 13 round-up of the Artificial Analysis Intelligence Index lists Muse Wonder Full at 56 and Muse Wonder at 52, in a field where the leading score is 58. That places the family firmly in the top group, which is a notable result for a first release at this scale.

For context on how quickly this line has moved, see our coverage of Muse Spark 1.3 and its benchmark results and the launch of Muse Code as Meta's coding agent.

What We Still Do Not Know

Several details are not yet public in readable form. Meta's post does not state a context window, a parameter count or whether weights will be released, so the models should be treated as API and subscription products for now. Meta says the models were trained at its gigawatt-scale Prometheus data center, and that a full safety report with additional evaluations will be available soon.

Meta also says Muse Wonder will soon power experiences in the Muse app, rolling out first to subscribers. No date is given.

Why It Matters

The interesting shift is where the competition is happening. For two years, model releases were compared mostly on reasoning and coding scores. Muse Wonder is pitched on remembering what matters to a user and staying on task, and it arrives with a benchmark designed to measure exactly that. If memory quality becomes a standard axis of evaluation, everyone building assistants benefits from clearer measurements.

I will revisit this once the safety report and independent memory evaluations are published. More model releases are in our AI coverage.

Sources: Meta, Muse Wonder announcement — October 12, 2026; TechCrunch — October 12, 2026; VentureBeat — October 13, 2026.