Meta on Monday, October 12, 2026, introduced the Muse Wonder series, made up of two models — Muse Wonder and Muse Wonder Full — that the company describes as the most capable it has built so far. According to Meta’s research blog, both deliver state-of-the-art results on personal and professional agentic tasks, memory and coding, and represent “the next step” on what the company calls its scaling ladder toward “personal superintelligence.” TechCrunch and Axios described it as Meta’s first serious attempt to compete at the frontier tier.
The launch comes fourteen months after Meta began a ground-up overhaul of its artificial intelligence work under Meta Superintelligence Labs, the division led by Scale AI founder Alexandr Wang. Until now, the company had released the lightweight Muse Spark model and the consumer Muse agent, but had avoided the kind of top-capability releases with which OpenAI, Anthropic and Google set the industry’s pace. Axios reported that the new models were known internally by the code name “Watermelon.”
Availability is staggered. Muse Wonder is usable from day one through the Meta Model API and Muse Code, the company’s coding environment; Muse Wonder Full, the more powerful variant, is reserved for Muse Code subscribers and, according to TechCrunch, will not be served through the API. Meta added that Wonder will soon power experiences inside the Muse assistant, again starting with subscribers, as well as artifact generation and other processes that need more reasoning capacity.
Pricing places the model at the high end of the market. On the API, Muse Wonder will cost $4 per million input tokens and $20 per million output tokens, a rate TechCrunch compares with OpenAI’s GPT-6 Sol or Anthropic’s Claude Opus 5.5. According to the tests Meta presents in its model card, Wonder and Wonder Full achieve scores comparable to Fable 5.1 and GPT-6 Astra on major benchmarks, though these are the company’s own measurements.
The centerpiece of the announcement is what Meta calls “Active Memory”: the model’s ability to carry important context about a person across conversations and proactively retrieve the details relevant to the task at hand. To back that claim, the company highlighted MemoryBench, a continual-learning and memory-construction benchmark on which, by its data, Muse Wonder beats Fable, Opus and Astra. “We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months and beyond,” research scientist Kavya Kopparapu said in a launch video.
The second concept is “Active Focus,” which describes how the model persists toward a hard goal: it tries alternative approaches, works through blockers, adapts to new information and, according to Meta, asks for clarification when needed and checks in before taking important actions. The company measures that behavior with FocusBench and sustained execution with SWE-Marathon, a test of long-running work on software projects. It also cites evaluations in law, finance and health, such as Harvey’s Legal Agent Benchmark, which requires working with source documents and following detailed instructions.
On coding, Meta says both models deliver frontier performance and were trained on diverse, long-running tasks to sustain progress through repeated cycles of testing and revision. Published demos include a 3D kitchen game generated from image inputs, a personalized running plan built by working directly in the browser, designing and ordering custom merchandise from photos on an Instagram profile, and producing a wedding seating chart under complex matching requirements.
Meta devotes a section to generalization, a sensitive point in an industry full of complaints about “benchmaxxing,” or tuning models to shine on specific tests. The company compares Muse Spark 1.3 with Muse Wonder Full and argues that, at equal performance on tasks targeted during post-training, Wonder Full does better on unseen tasks than Spark’s trend would predict. It attributes this to the new models being built on larger base models, whose post-training gains transfer better to unforeseen situations.
The models were trained at Prometheus, the gigawatt-scale data center Meta brought online this year in central Ohio. That detail ties the announcement to the infrastructure race: big tech companies are investing tens of billions of dollars in compute, and investor pressure to show tangible results has grown. For Meta, a competitive in-house model reduces reliance on third parties and lets it build advanced capabilities into products used by billions of people.
On safety, Meta says Muse Wonder offers its strongest safeguards to date against chemical, biological, radiological and nuclear misuse, as well as cyber misuse, while preserving access to defensive capabilities under its Meta Superintelligence Scaling Framework. The company calls it its “most aligned” model and promises a full report with additional safety evaluations soon. Axios noted that a model card with some details was due on Monday, with a fuller report to follow.
Context makes that promise matter. In recent weeks, several labs have reported agents that bypassed restrictions in test environments, and European regulators have opened inquiries into loss-of-control incidents. A model designed to remember personal information for months and act on a user’s behalf across browsers and apps also raises obvious privacy questions. Meta says it trained the Wonder family to exercise discretion with private information and resist adversarial instructions, but those claims still need independent verification.
“These are true frontier models, competitive with the best models in the world, and are particularly state of the art on personal and professional agentic work,” Alexandr Wang told Axios. Mark Zuckerberg, for his part, wrote that Muse Wonder is “a much larger and more advanced class of models that brings us to the frontier.” Euronews reported that Meta also leaned on data from the independent firm Artificial Analysis to support its comparison with models from Anthropic, OpenAI and Google.
What remains to be seen is whether Meta’s claimed advantages hold up in independent testing and how quickly the technology reaches the company’s apps. If active memory works as described, Meta could stand out in long-term personal assistants, a field where its huge user base is an advantage; if not, Muse Wonder will be another powerful model in an increasingly crowded market with prices under pressure. For now, the launch confirms that the frontier race is no longer a three-lab contest.
