Meta enters the frontier with Muse Wonder: the data comes from Artificial Analysis, the challenge is sustaining pace and trust

By Mashable · October 12, 2026.

Meta has launched Muse Wonder, a new family of AI models designed for agentic tasks, meaning those in which the model plans and executes steps on its own rather than simply answering. According to Mashable, it is already reaching developers through the Meta Model API and Muse Code. The most powerful variant, Muse Wonder Full, is available to Muse Code subscribers.

The key data point comes from a third party, not Meta. According to Artificial Analysis, an independent evaluation organization, Muse Wonder Full ranks second on its Intelligence Index (a composite index that weights several tests), behind only Claude Opus 5.5 Max and tied with Claude Sonnet 5.5 Max. It sits ahead of GPT-Astra, Fable 5 and the newly launched Gemini 4, again according to that index. Artificial Analysis adds that Muse Wonder Full is in the top three on six of the index's ten evaluations, with particular strength in agentic knowledge work and scientific reasoning.

The launch comes six weeks after Muse Spark 1.3, the latest version of a model family that debuted in April. Meta is also presenting its own internal results and a model card, and says a full safety evaluation report will be available soon. As always, it is worth separating what a third party measures from what the maker publishes.

The model includes two features Meta calls Active Focus (persisting on a difficult task, trying other approaches, recovering from errors and checking in before important actions) and Active Memory (remembering what is relevant across conversations). Meta also says Muse Wonder will avoid sharing private information without permission. According to Mashable, this has been a flagged problem with the Muse agent; Meta itself does not describe it that way in the material cited.

As for public access, according to CNBC as cited by Mashable, the new models will come to Muse for those paying $20 or $100 a month. In other words, Meta's frontier will not be free.

**Our take.** Thesis: what matters most is not the exact ranking position, but that the frontier is no longer a club of two or three. In our quality record, Anthropic and OpenAI led, with China close behind and Google trailing; Meta appearing second on an independent index confirms that the leaders' edge is narrowing and that the model on its own is becoming commoditized. In general, when several players reach similar capabilities, value shifts to distribution, integration and trust.

This is where Meta has both its biggest advantage and its biggest burden. It has massive distribution and an agent, Muse, which according to Mashable has surpassed 5 million downloads. But an agent with persistent memory and the ability to act demands a spotless privacy record, and the episodes Mashable reports (reports that Muse stores files of people close to the user, or that it sent a Marketplace buyer to a user's home) are exactly the kind of failure that erodes that trust. These are press reports, not facts we can treat as proven, but they show where the bar is.

There are caveats. A single composite index is not reality: we have previously warned that indexes can favor generous configuration versions (here, "Max" variants), and Meta's own results are marketing until they are replicated. The figure to watch will be consistency on real long-horizon tasks, not the podium position, along with the still-pending safety report.

For readers, the underlying good news is competition: more labs at the frontier mean pressure on prices and on the pace of improvement, ingredients of long-term abundance. In the short term, the trade-off is more autonomous agents with access to personal data and governance that lags behind. The control point, once again, is shifting from the model to what the agent can see and decide.