Meta launches Muse Wonder frontier models with a focus on personal agents

What Meta announced

Meta has introduced the Muse Wonder series, two models the company describes as its most capable yet for personal and professional agentic tasks. The lineup consists of Muse Wonder, available through the Meta Model API and in Muse Code, and the larger Muse Wonder Full, which is restricted to Muse Code subscribers.

According to The Verge, the launch extends a fast-moving run for Meta's revamped Superintelligence Labs division, which began with the Muse Spark model and the Muse agent. Meta says Wonder will eventually power the Muse assistant, reaching subscribers first, and that specialized models derived from it will follow for all Muse users — though The Verge notes the company provided no timeline or specifics for either step.

In its blog post, Meta frames the release as the next stage of a deliberate scaling program aimed at what it calls personal superintelligence, and says the models were trained at its gigawatt-scale Prometheus data center after a ground-up rebuild of its AI stack that began fourteen months ago.

Memory, focus and sustained coding

The post highlights two capabilities central to the release. Active Memory is described as carrying context about a user across conversations and actively retrieving the details relevant to the current task. Active Focus covers sustained effort toward a difficult goal: trying alternative approaches, working around blockers, asking for clarification when needed and checking in before taking important actions.

Meta says it measures these qualities with FocusBench and with SWE-Marathon, a benchmark for long-running software projects, and claims frontier-level coding performance, demonstrated in a demo where Wonder builds a playable 3D kitchen game from customizable image inputs. Domain expertise is evaluated in law, finance and health, including Harvey's Legal Agent Benchmark, which requires working from source documents to produce legal work products — the kind of task involved in understanding a lease or interpreting financial and health information.

The claims, and who has verified them

Meta reports state-of-the-art results across benchmarks for agents, memory and coding. The Verge reports that Meta positions the new models as rivals to OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1, based on internal testing — and notes that it remains unclear how they stack up against Google's Gemini 4 Argon, whose capabilities are still emerging.

Meta also addresses a familiar line of skepticism directly. Acknowledging complaints about "benchmaxxing" — models tuned to shine on popular benchmarks — the company says it studied how performance on tasks targeted during post-training relates to performance on unfamiliar tasks as test-time compute grows. It reports that Muse Wonder Full does better on out-of-distribution tasks than the trend set by the earlier Muse Spark 1.3 would predict, a gap Meta attributes to a stronger base model. That breadth matters, the company argues, because a personal agent faces combinations of goals, tools and constraints that cannot all be anticipated during training.

Safety commitments

Because a personal agent reads messages, calendars and private documents and acts on a user's behalf, Meta says it trained the Wonder family to handle private information with care and to resist adversarial instructions. The company claims its strongest safeguards to date against chemical, biological, radiological, nuclear and cyber misuse, while preserving frontier cyber capabilities for defensive use under its Meta Superintelligence Scaling Framework. A full report with additional safety evaluations is promised but has not yet been published.

Why it matters

Meta is staking its AI strategy on agents that remember, plan and persist, rather than on a chatbot interface alone, and Muse Wonder is the model meant to make that real. For now, every headline claim — the benchmark wins, the memory and focus behavior, the safety posture — comes from Meta itself: the results are from internal testing, the safety report is pending, and there is no announced schedule for when Wonder reaches the billions of people using Meta's products. If the claims hold up outside Meta's labs, the gap with OpenAI, Anthropic and Google narrows considerably; if they do not, Wonder joins the long list of frontier launches whose benchmark scores outpaced their everyday usefulness. The subscriber-first rollout also signals how Meta intends to convert its heavy infrastructure spending into recurring revenue.