Meta ships Muse Wonder with no benchmarks or prices

Meta says it has a frontier model. It has not shown you the numbers.

Meta released two models on Oct. 12: Muse Wonder and Muse Wonder Full. The company calls them its "most capable models to date." Muse Wonder is available now to developers through the Meta Model API and Muse Code. Muse Wonder Full is available now to Muse Code subscribers.

That is a shipped product you can call today. It is also a launch with gaps.

Meta built Muse Wonder for agents

Meta says the models aim for "state-of-the-art performance on personal and professional agentic tasks." Its examples are ordinary chores: working through an insurance claim, understanding a lease, making sense of someone's finances, building a working app. It names law, finance and health as expert domains and claims "frontier coding performance."

Two features carry the pitch. Active Memory, Meta says, carries what matters to you across conversations and brings back details that fit the current task. Active Focus keeps the model working toward a goal when a task gets hard: it tries other approaches, recovers from errors, asks for clarification when it needs to and checks in before following through on anything important.

Muse Wonder builds on Muse Spark, Meta's earlier model. Meta says it will soon power its Muse app, rolling out to subscribers first.

The launch leaves out what you would use to decide

Meta's announcement has no benchmark scores. It has no price for the Meta Model API. It gives no context window and no parameter count.

Meta's research post names some of the tests it ran, including Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon. The results sit in a chart against "other frontier models" the text does not name. The only model the text names for comparison is Meta's own Muse Spark 1.3.

The safety picture is pending too. Meta says it trained the models to avoid sharing private information without permission, to decline requests to help with harmful acts and to strengthen defenses against cyberattacks, while still letting security teams find flaws in their own code. It says it released them under its Meta Superintelligence Scaling Framework. A full report with more safety evaluations "will be available soon," according to the company.

Yes, every lab launches with a chart. Most also launch with a system card and a price list. Meta shipped the model first and the paperwork later.

What you can do with it this week

If you build AI agents, Muse Wonder is a new option on a real API. That alone earns it a test.

Pick one multi-step task you already run on another model. Choose the kind Meta is pitching: a long document to read and act on, a coding job that takes more than one pass, a workflow where the model has to recover from its own mistake. Run it on both. Count the retries. Add the cost once Meta publishes prices.

Keep sensitive data out until the safety report is public. Active Memory is built to remember across conversations, so read Meta's terms on what it keeps before you hand it a client's finances.

If you only use AI in a chat app, there is nothing to change today. Muse Wonder reaches Meta's Muse app soon, subscribers first.

Watch the safety report and the price sheet

Two documents decide whether this launch matters to you: the safety report Meta promised and the Meta Model API price list.

The first tells you what Meta found when it tried to break the model. The second tells you whether switching pays.

Until both land, Meta has made a claim. Test it.