Just over a year after forming Meta Superintelligence Labs, the company moved from its earlier Muse Spark models to what it calls a far larger class of AI. On Monday, Meta launched Muse Wonder and the more capable Muse Wonder Full, saying the models can compete with OpenAI and Anthropic at the technological frontier.¹

Mark Zuckerberg framed the release as the payoff from that push, writing that Muse Wonder is a “much larger and more advanced class of models that brings us to the frontier.”Mark Zuckerberg (finkd)² Meta is offering Wonder through its Model API and Muse Code, while Wonder Full is available to Muse Code subscribers. It also plans to make Wonder the engine for future versions of its Muse assistant, initially for paying users.¹

The company’s case rests on agentic work: AI that can retain user information, stay on task and adapt when an approach fails. Meta says its “Active Memory” can carry relevant details across conversations, while “Active Focus” helps an agent pursue longer, multi-step jobs—such as an insurance dispute—rather than simply answer a prompt.³

Meta’s internal benchmarks say Wonder matches or surpasses leading OpenAI and Anthropic systems in coding, professional work and personal-agent tasks. Alexandr Wang, who leads Meta Superintelligence Labs, called them “true frontier models” and said memory is essential because users should not have to repeatedly explain themselves.¹ Aravind Srinivas offered an early note of support, saying the internally named Watermelon model compared favorably with Fable 5.1 on WANDR evaluations.Aravind Srinivas (AravSrinivas)⁴

But the competitive picture is not settled. Meta has not detailed how Wonder compares with Google’s Gemini 4 Argon, and the strongest evidence so far comes from Meta’s own testing.⁵ The larger challenge is whether third-party evaluations back the company’s claims—and whether agents built to remember personal information can reach billions of users without creating fresh privacy and security risks.¹