Meta introduces Muse Wonder, its answer for agentic frontier AI

Alexandr Wang, Meta's chief AI officer, in a promotional video for Muse Wonder. Meta


There have been two running questions for ever-pivoting Facebook parent Meta about its turbulent AI efforts. 

First: Will the 2025 revamp of its Superintelligence Labs group result in more competitive AI output? 

Second: Will its $145 billion capex spending spree pay off?

Meta spent much of this year trying to tackle the first question. The company first captured everyone’s attention with the April launch of its Muse Spark family of LLMs, then last month with the launch of its personal AI agent Muse. 

On Monday, the company doubled down with the launch of Muse Wonder and Muse Wonder Full, which it calls its frontier models for agents.

Now hang on, you’ll say: Isn’t Muse Spark supposed to be agentic? It is—but Meta’s new Muse Wonder duo is bigger, smarter, and by extension slower than its sparky sibling. 

That’s what frontier AI models are all about, and that’s why Muse Wonder was trained at Meta's gigawatt-scale Prometheus data center. (And yes, it will power some experiences inside the Muse agent, as well as develop future specialized AI models.)

In a blog post detailing various benchmarks, Meta says its new models can hang with frontier peers like OpenAI’s GPT-6 Astra as well as Anthropic’s Claude Fable 5.1 and Opus 5.5.

The company also touts two new capabilities of the models: Active Memory, which it describes as the ability to carry context across conversations, and Active Focus, which it describes as a kind of resilience when initial approaches don’t achieve the task at hand.

And what of that pesky second question about ROI? Unlike peers, Meta has predominantly dedicated its AI efforts to working within its consumer-focused platform ecosystem. While its agents are unlikely to process your department’s expense reports or redline corporate contracts—chances are high your employer prefers an enterprise-ready alternative—they are more than happy to navigate your insurance claim, understand your apartment lease, or build you a mobile app for your small business.

All of that helps Meta diversify its revenue away from the digital advertising that currently accounts for upwards of 98% of Meta’s $201 billion in annual revenue.

It will take an awful lot of Muse subscribers to move that needle, but it’s not impossible. Deutsche Bank analyst Benjamin Black mused only last week that Meta’s personal AI agent could conceivably represent up to 8% of Meta's revenue by 2030, or about $36 billion in sales. Is it enough to offset all those AI engineers’ salaries? Not a chance. But no one knows better than Meta that the status quo won’t cut it. —Andrew Nusca