Meta released Muse Wonder and Muse Wonder Full on October 12, describing them as the company’s strongest models to date and its first family of frontier systems. The pair is built on a base model the company calls Muse Spark, and both are aimed at agent work rather than conversation.
The tasks Meta used to describe the release are drawn from office life. The company said the models handle insurance claims, reading a lease, and building an application that runs. Those are jobs with a document at one end and a decision at the other, and they require a model to read a contract, hold a set of constraints in view and produce something a person can use.
Access is split along a line the industry has been drawing for a year. Muse Wonder became available on the day of the announcement through the Meta Model API and through Muse Code for developers. Muse Wonder Full is reserved for subscribers to Muse Code, the company’s coding product. The Muse application sells subscriptions at two prices, $20 and $100 a month.
Putting the stronger model behind the higher price band is a change from how consumer AI products were first sold. The early pattern was to give the best available model to everyone and charge a single fee. The newer pattern charges by the difficulty of the work, on the theory that a model that reads a lease and returns a decision is worth more to a business than a model that drafts an email.
The person leading the effort is Alexandr Wang, according to the company. Meta has spent the past two years rebuilding its AI organization around a smaller number of senior researchers, and the release gives that group its first model family to put in front of paying customers.
Developers are the route to distribution, and Meta’s terms reflect it. Opening Muse Wonder through an API lets other companies build products on the model without adopting Meta’s own applications. Reserving the Full version for Muse Code subscribers gives the company a reason for developers to pay it directly rather than consume the model through an intermediary.
Splitting a model family in two also solves a production problem. A single system that answers questions and one that completes multi-step tasks draw on the same weights, and offering both at one price forces the operator to serve the heaviest requests at the lightest tier’s rate. Meta’s split lets it meter the harder work separately, and it gives the company a way to raise what it charges without changing what a casual user pays.
Enterprise sales look different from consumer ones, and Meta has less practice at them. Companies buying agent products want contracts, data handling terms and support commitments, none of which a $20 monthly subscription provides. Selling through an API to other developers lets Meta reach those buyers indirectly, since the companies that adopt Muse Wonder may already sell software to the businesses Meta would otherwise have to court.
Analysts said the pricing structure is a test of whether agent work can support higher subscription prices than chat. The $100 tier is the highest price Meta has attached to a subscription for its models, and it arrives at a moment when every large model maker is looking for a revenue model that covers inference costs rather than simply attracting users.
Those costs are the constraint behind the tiering. Serving an agent that reads a twenty-page lease consumes far more computing than answering a question, and the difference lands on the operator. Charging a flat fee for unlimited use of a heavy model is the fastest way to lose money on the customers who like the product most, which is why the industry has moved toward usage limits and price bands.
The competitive position is crowded. Google, OpenAI and Anthropic all sell models for the same document-and-decision work, and all three have enterprise relationships Meta has to win. Meta’s advantage is its distribution: it operates the applications where billions of people already spend their time, and it can put a model in front of users without asking anyone to download anything new.
What the company has not described is how the models were trained, how large they are, or what they score on standard tests. Meta’s release materials describe capability by task rather than by benchmark, a shift several labs have made as public benchmarks have become less informative and as disclosure has become a competitive decision.
Enterprise buying behavior will decide whether the tier structure holds. Companies evaluating agent products generally test them on their own documents, then renegotiate on volume. A model priced at $100 a month per seat invites exactly that conversation, and the answer depends on whether the work the model does replaces labor or merely assists it.
Meta has not said what happens to models it released under earlier names, or whether the Muse family replaces them. It also has not described data-handling terms for enterprise customers, a question that decides most large contracts in this category and one the company has answered cautiously in the past.
For Meta, the release is the first commercial test of an organization it rebuilt from the top down. The company has spent heavily on computing and on researchers, and the return now depends on whether businesses will pay for agents that do office work. The two price bands are the first measure of how much anyone will pay.
