Meta just built AI models that rival anything OpenAI or Anthropic can produce — and it's keeping the best one locked away from the public.

The company released Muse Wonder and Muse Wonder Full on Monday, its first true frontier AI models. But while the standard Wonder model is available through a commercial API, the full-power Wonder Full is reserved for Muse Code subscribers paying $50 to $200 a month. Neither model ships with publicly downloadable weights, marking a sharp break from Meta's earlier Llama models, which helped popularize open-access AI. The company that once let the public run powerful AI on their own hardware now wants to be the gatekeeper.

According to independent benchmarking organization Artificial Analysis, Wonder Full scored a 56 on its Intelligence Index — tying Anthropic's Claude Sonnet 5.5 and beating OpenAI's GPT-6 Astra, Google's Gemini 4 Argon, and Anthropic's Claude Fable 5.1, which each scored 53. The standard Muse Wonder earned a 52. VentureBeat reported that only Anthropic's flagship Claude Opus 5.5, at 58, scores higher among these models.

So what does the public get? Access to the weaker model at $4 per million input tokens and $20 per million output tokens — pricing TechCrunch noted is roughly equivalent to GPT-6 Sol or Opus 5.5. The real firepower stays inside Meta's walled garden.

This is a deliberate strategy. Engadget reported that Meta expects to spend as much as $145 billion this year on its AI plans. CEO Mark Zuckerberg told the Financial Times that reaching this point "required a large amount of compute and investment, and a lot of very rigorous scientific work." That investment buys leverage — and Meta is using it to funnel users into its own subscription ecosystem rather than releasing weights the way it did with Llama.

Both models feature what Meta calls "Active Memory" — the ability to carry personal and professional context across conversations — and "Active Focus," a looping system that keeps the AI working toward a goal even when initial attempts fail. Digital Trends framed these capabilities as user-friendly upgrades. But they also mean Meta's AI remembers everything about you and persists in its tasks — all inside a closed system the company controls end to end.

The release comes roughly 14 months after Zuckerberg restructured Meta's AI efforts into Meta Superintelligence Labs, recruiting former Scale AI CEO Alexandr Wang. TechCrunch noted this is the first frontier model release since that reorganization. The shift from open weights to proprietary access isn't an accident — it's the point.

VentureBeat was the only outlet to flag the obvious tension: Meta's earlier Llama models "helped popularize downloadable AI models that organizations could run on their own infrastructure," and now neither Wonder model offers that. Digital Trends, Engadget, and TechCrunch all soft-pedaled the closure, treating the subscription lockup as a pricing detail rather than a fundamental change in who gets to control powerful tools.

Meta says it trained the Wonder family to "avoid sharing private information without permission" and "turn down requests to assist in performing actually harmful acts," per Digital Trends. Translation: Meta decides what's harmful, Meta decides what you're allowed to do, and Meta keeps the most capable version of the tool out of your hands entirely.

The question isn't whether Meta built something impressive. It did. The question is why Americans should trust a single corporation to decide who gets access to the most powerful tools of the next decade — especially when that corporation used open-source goodwill to build its position, then slammed the door behind it.