Meta Launches Muse Wonder: An Agent Model With Active Memory and Active Focus

Meta announced Muse Wonder on October 12, and for a model family named after wonder, the substance is procedural: two named capabilities, Active Memory and Active Focus, that turn Muse from a chat assistant into an agent that remembers you across conversations and sustains effort on hard goals. Muse Wonder is available today via the Meta Model API and Muse Code; the stronger Muse Wonder Full ships to Muse Code subscribers. It was trained at Meta’s gigawatt-scale Prometheus datacenter, and Meta claims state-of-the-art results on agentic, memory, and coding benchmarks, evaluated through two new benchmarks, FocusBench and SWE-Marathon.

The Two Named Capabilities Are the Actual Product

Active Memory carries context about the user across conversations and retrieves relevant details on demand, which puts Muse in direct competition with every memory feature shipped this year, including Claude’s cross-surface memory. Active Focus is the more interesting claim: sustained effort on hard goals, trying alternatives, asking clarifying questions, and checking in before major actions. That last clause reads as a direct answer to the month’s agent-safety arc, in which Meta’s Muse was found reading messages with permissions off and Muse’s zero-day exposed its auth token. A model that checks in before major actions is a design commitment to exactly the failure mode Muse already demonstrated. Whether it holds under adversarial testing is the open question, but the capability is at least named as a first-class feature rather than a permission toggle.

The Benchmark Strategy: Two New Benchmarks

Meta evaluated Muse Wonder on FocusBench (focus) and SWE-Marathon (long-running software projects), both new, which is the now-standard move of launching a model with benchmarks built for its strengths. The demos skew agentic and consumer-adjacent: a Saturday run-planning agent working in-browser across apps with OpenStreetMap data, custom swag designed and ordered from personal Instagram photos, and a wedding seating chart generated from complex matching constraints. Domain claims reach further: legal work benchmarked against Harvey’s Legal Agent Benchmark, frontier coding trained on long-running software tasks, and a demo building a playable 3D kitchen game from image inputs. A generalization analysis claims Muse Wonder Full outperforms the Muse Spark 1.3 trend on out-of-distribution tasks at equivalent in-distribution scores, attributed to stronger base models.

The Context: Muse’s Troubled Month, and Prometheus

Muse’s month has been the worst of any consumer agent: the auth-token zero-day, the permissions bypass, Amazon blocking it from shopping, and an NYT inside account of Zuckerberg pulling the trigger despite safety concerns. Muse Wonder is the response: a model trained at the gigawatt-scale Prometheus datacenter on the Muse Spark scaling approach, with safety claims of privacy discretion, adversarial-instruction robustness, and the strongest CBRN and cyber safeguards to date, with a full safety report promised. The market test is immediate: everything Meta promises about Active Focus and check-ins before major actions will be evaluated against four weeks of documented Muse failures.

What to Watch

Three things. First, FocusBench and SWE-Marathon methodology: new vendor-built benchmarks evaluated at launch are marketing until independent replication, and the memory claims in particular need adversarial testing. Second, the Muse app integration: Muse Wonder will soon power the Muse app for subscribers first, which puts Active Memory inside the same app whose permissions bypass made October’s headlines; how Meta scopes memory access there is the real test of its privacy discretion claims. Third, the competitive math: Meta now has an agent model family trained at gigawatt scale, OpenAI has Dots with purchase authority, and Anthropic has the 5.5 family with the price ladder. The agent market has its three-way race, and October was the month every contender showed its hand.