Meta Pushes Into the AI Frontier With Muse Wonder, Its Most Capable Models Yet

Meta's Superintelligence Labs unveiled Muse Wonder and Muse Wonder Full on Monday, two proprietary, agent-focused models it says can trade blows with OpenAI's and Anthropic's best — but only the weaker of the two is available through a public API.

Meta on Monday introduced Muse Wonder and Muse Wonder Full, calling them its most capable models yet and its first real entries in the frontier race that OpenAI, Anthropic and Google have led. They come from Meta Superintelligence Labs, the group led by former Scale AI chief Alexandr Wang, and they are built less for chat than for agents — software that can work through an insurance claim, parse a lease, dig into someone's finances, or build and test a working app.

The release marks a sharp turn for Meta. Its earlier Llama models were celebrated for being downloadable and free to run on your own hardware. Neither Wonder model will be: both are proprietary, and neither is offered with publicly downloadable weights. Only the weaker of the two is open to outside developers today. Muse Wonder is sold through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20. Meta's strongest model, Muse Wonder Full, is kept behind Muse Code, its coding assistant, reachable only through $50- and $200-a-month subscription tiers — not a pay-as-you-go API enterprises can build on.

On the numbers, Meta is now a genuine contender. The independent benchmarking outfit Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index, tying Anthropic's Claude Sonnet 5.5 and edging past OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1, each at 53. Only Anthropic's flagship Claude Opus 5.5 scores higher, at 58. The cheaper standard Muse Wonder lands at 52. Meta's own model card shows Wonder Full leading on long-running software engineering and legal tasks, though Opus 5.5 still wins on the aggregate measure and on health questions.

What Meta is really selling is endurance. The models accept text, images, audio and video, carry a context window of roughly one million tokens, and add two features aimed at long jobs: Active Memory, which carries context across conversations and pulls back what is relevant, and Active Focus, which keeps an agent pursuing a goal when it gets hard — trying other approaches, recovering from errors and checking in before anything consequential. Zuckerberg described Wonder as a much larger and more advanced class of model than Meta's earlier Muse Spark family, and Meta says Wonder will begin powering the Muse assistant itself, starting with subscribers.

The fine print matters for anyone tempted to deploy it. Muse Code subscriptions default to Contributor mode, which lets Meta use submitted prompts, code and outputs to improve future models; the pay-as-you-go API runs in a standard mode where Meta says inputs are not used for training. Meta's shares rose on the news. Yet the most capable version of Meta's new frontier remains something you can rent inside Meta's own tool — not something you can own, host or wire into a rival's product. On the evidence so far, Meta has arrived at the frontier. It has not yet moved in.

Sources: Meta Newsroom, Meta Research, TechCrunch, CNET, Mashable