Meta Calls Muse Wonder Its First Frontier Models — Without Independent Testing
On Monday, October 12, 2026, Meta launched two AI models, Muse Wonder and Muse Wonder Full, the first the company itself places in the frontier class. The models became available the same day via the Meta Model API and Muse Code, but the benchmark figures and the frontier claim are so far based solely on data supplied by Meta itself — a model card and a full safety report are still missing.
What Has Actually Been Launched
On Monday, Meta launched a new series of advanced flagship models in an attempt to catch up with competitors such as Anthropic and OpenAI, according to The New York Times (a432c7a8).
There are two models in the family:
- Muse Wonder is the standard model. According to the NYT, it will soon power parts of Muse, Meta's agent app (a432c7a8).
- Muse Wonder Full is the most powerful model and is reserved for users who subscribe to Meta's premium coding tool Muse Code (a432c7a8).
Both models became available the same day through the Meta Model API and Muse Code. The market reacted quickly: Meta's stock traded 3.7 percent higher on Monday following the announcement, according to Investing.com via Yahoo (fc0185ec). This figure has not yet been confirmed by other sources.
The Background: Watermelon, a July Pause, and Wang's Division
The Wonder family is the second model series developed under Meta's Superintelligence Labs. Meta began the work after Alexandr Wang, founder of the AI company Scale AI, was hired as chief AI officer in June 2025, according to the NYT (a432c7a8).
Internally, the project went under the codename Watermelon, the NYT reports (a432c7a8). In July, Meta paused a phase of model development known as "pre-training," which delayed the launch until this month, the newspaper has previously reported (a432c7a8).
Mark Zuckerberg confirmed the launch on X, drawing the long arc. "It's been just over a year since we started Meta Superintelligence Labs," he wrote, according to Fox News Digital. "Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier." He also stated that the models were trained at the Prometheus data center in Ohio (6d099145).
What Meta Claims — and What Is Not Verified
So far, all performance claims come from the company itself.
According to data supplied by Meta, the two new models score nearly as well as Anthropic's most advanced models on tests measuring reasoning, writing, and coding (a432c7a8). These are thus company figures relayed through secondary sources — no independent testing exists yet.
Wang pushed the frontier claim hard in his interview with Axios: "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," he said (a018b45a).
Active Memory and Active Focus
Meta's two main features in the launch concern agentic work over time, Axios writes:
- Active Memory is intended to let the model retain information about the user across conversations.
- Active Focus is intended to enable long-term, goal-directed agentic work.
Together, the features are meant to help AI systems perform more useful work over longer periods — a claim from Meta that Axios relays without independent documentation (a018b45a).
Safety, Pace, and Open Source
Meta says the models were developed according to the company's Meta Superintelligence Scaling Framework, which is supposed to ensure that the models do not share private information without permission and refuse harmful requests (fc0185ec). But the full safety report is not available, making it impossible to assess how the framework actually works in detail.
Zuckerberg, in an interview with the Financial Times — relayed via Investing.com/Yahoo — said Meta has reached the AI frontier and caught up with its competitors. He rejected calls to slow the pace of development and signaled that the company will keep investing because it is "already working on the next step up" (fc0185ec).
The open-source question — historically one of Meta's hallmarks in the field — remains unresolved. Meta has not said whether any of the Wonder models will be opened. Wang pointed out that Meta recently open-sourced the weights for Muse Glitter, along with the technology needed to run the Muse agent on hardware, but he said nothing concrete about the Wonder family (a018b45a).
What Remains
Several key pieces are still missing from any assessment of the Wonder family:
- The model card was planned to arrive on Monday but was not present in the source material at launch.
- The fuller safety report is still pending, and without it, the claims about protection against misuse cannot be verified.
- Independent testing is lacking. As long as the benchmark figures come from Meta itself, the frontier claim — and the assertion of near parity with Anthropic — is a company statement, not a verified conclusion.
The conclusion for now: Meta has launched two new flagship models with clear ambitions for the frontier position, and the company's leadership presents them as a breakthrough. Whether they actually are depends on the documentation that remains outstanding — and on the independent testing that has not yet begun.