SAN FRANCISCO, CALIFORNIA – Meta just leapfrogged to the lead in the world’s most important race.
On October 12, the company introduced Muse Wonder and Muse Wonder Full, its most capable AI models yet. Meta says they deliver leading performance on tasks that increasingly matter to AI users: remembering context, working through complicated projects, and getting useful work done.
The release brings together several pieces of Meta’s strategy: a personal agent, a growing wearable lineup, and increasingly capable models that can power both.
The comeback has substance
Meta describes the new models as the product of a ground-up overhaul of its AI efforts that began 14 months ago. Muse Wonder builds on Muse Spark, with training at the company’s gigawatt-scale Prometheus data center.
The results deserve attention, with some qualifications.
In its published model card, Meta reports that Muse Wonder Full scores 59.1 on its Personal Superintelligence Index, ahead of the listed OpenAI and Anthropic models. That index combines seven evaluations of personal and professional work.
But it is Meta’s internal composite. It isn’t a universal measure of intelligence.
The same table shows Wonder Full ahead on several coding tests, while competitors lead elsewhere. On the Artificial Analysis Index, for example, Wonder Full scores 56, below Claude Opus 5.5’s 58.
Meta also acknowledges that its internally run competitor tests may not optimally tune prompts, tools, and runtime for proprietary models.
Remembering and following through
Meta emphasizes two capabilities: Active Memory and Active Focus.
Active Memory is meant to carry useful context across conversations and retrieve it when it becomes relevant. Active Focus is meant to keep the agent working through obstacles, trying another approach and adapting to new information.
Meta says the models ask for clarification and check in before important actions. Its demonstrations include planning a run around personal circumstances and preparing a wedding seating chart with complicated requirements.
These capabilities address a familiar frustration. An assistant can produce a convincing answer, then lose track of a requirement or stop when the next step gets difficult.
Imagine explaining your family’s scheduling constraints once, then having an assistant apply them when a new appointment arrives. Or asking it to resolve a software problem and having it work through revisions until the project functions.
Meta already has somewhere to put it
In September, Meta launched Muse, a personal agent designed to work through its own dedicated virtual computer and browser. Users can communicate with it in the Muse app or through WhatsApp.
Meta says it can handle tasks such as sending email and booking travel, with permission checks for sensitive actions.
That gives the new models a product to improve.
For now, access has limits. Muse Wonder is available through Meta’s developer API and Muse Code. Wonder Full is available to Muse Code subscribers. Meta says Wonder will reach Muse experiences soon, beginning with subscribers. A launch announcement should not be confused with a completed consumer rollout.
For developers, Meta’s launch guide also describes compatibility with existing OpenAI-style clients, reducing the work needed to try its API. Wonder Full is offered through Muse Code’s Power and Maximum Usage plans.
The glasses add another opening
At its September 23 Connect event, Meta announced plans to bring Muse to its AI glasses.
It also introduced camera-free Ray-Ban Meta Audio glasses. Meta says those frames weigh 43 grams and offer up to 12 hours of use on a charge.
Glasses give an assistant another way into daily routines: a spoken request while walking, handling chores, or moving between appointments.
The strategic appeal is clear. Meta can compete over the intelligence of an assistant and the places people access it.
Whether users find that useful enough to become a habit remains an open question.
The advantage comes with a bill
Meta’s second-quarter results show the scale behind the strategy: 3.60 billion daily users across its family of apps in June, and $60.80 billion in quarterly revenue, up 28%.
Those are app users, not Muse users. Still, they represent an extraordinary potential distribution channel.
The spending is extraordinary, too. Meta guided to $130 billion-$145 billion in 2026 capital expenditures. Quarterly operating income fell 8%, and free cash flow was $784 million. Its ability to fund AI does not make the investment painless.
Trust presents another test. Meta says Muse conversations and data in users’ virtual machines are kept out of its advertising systems. It also allows users to disconnect services and opt out of using interactions for model training.
An assistant that remembers more and acts more needs users willing to grant it access.
Meta now has a stronger case on capability, a route into everyday life, and a business large enough to support the effort. The next test is whether those pieces produce dependable help that people choose to use again.
Tom’s Take
It’s easy to forget that consumer AI is still a very young technology, despite all the hype and impact.
It’s also easy to forget that in many cases, the person or company that invents a new technology isn’t the entity that makes it profitable and powerful.
Edison didn’t actually invent the lightbulb — he just commercialized it. Kodak invented the digital camera, and then sat on it for decades. We all know how that went for them!
The same dynamics could well end up applying to AI. Just because OpenAI and Anthropic pioneered consumer AI, that doesn’t mean they’ll be the ones to make it a viable business long term.
Meta has been quietly sitting on the sidelines, presumably learning a ton — for free — from the successes and failures of its competitors.
Now, it can finally strike, joining the market with a model that beats its younger competitors on multiple metrics — as well as the gigantic user base, cash flow, and data centers needed to scale it fast.
A few weeks ago, the AI race had three horses. Now it has four.
A version of this story originally appeared in The Generator on Medium.
