Skip to main content
Product2026-10-13•8 min read

Muse Wonder 1.0 Is Now on Multos — Meta's Frontier Model That Remembers

Active Memory and Active Focus are the two ideas behind Muse Wonder, and they are the right ones for agents that run for hours. Trained at Meta's gigawatt-scale Prometheus data center. Live on Multos Elite today.

O
Oghenekaro Oboido

Muse Wonder 1.0 is Meta's frontier model and it is on Multos Elite. It costs $4 per million input and $20 per million output, carries a 1M context window, and was trained at Meta's gigawatt-scale Prometheus datacenter. What makes it different from every other frontier model is not a benchmark score. It is that it was built to remember.

Active Memory: The Problem Nobody Solved

Most models forget. You start a session, explain your codebase, get somewhere, and the next session starts from zero. Meta's answer is Active Memory: carrying important context about you across conversations and actively retrieving the details relevant to the current task, rather than waiting for you to re-explain. For anyone building on Multos this maps onto a real problem. A project that spans weeks. A codebase you return to. A set of decisions you made three sessions ago and cannot quite recall. Wonder is the first model in this catalog that treats that continuity as a design goal rather than an accident of context length.

What That Looks Like Across Builds

The builds that benefit are the long ones. A refactor that touches forty files and cannot finish in one sitting. A product you are iterating on over months, where the requirements keep moving. A support tool that needs to remember a customer's history. A research app that accumulates findings. A migration that has to hold the old schema and the new one in mind at once. Wonder is less interesting for a weekend build and more interesting for the project that has been open in a tab for a month. If your build has a memory problem, this is the model aimed at it.

Active Focus: Working Through Blockers

The second half of Meta's pitch is Active Focus: continuing toward a goal by trying alternative approaches, working through blockers, and adapting to new information, while asking for clarification when needed and checking in before important actions. They evaluate it with FocusBench and SWE-Marathon, the latter testing work on long-running software projects. This matters for agentic builds specifically. A model that gives up at the first failing test is not useful for a build that runs unattended. A model that keeps trying, and knows when to stop and ask, is.

The Generalization Argument

Meta makes an argument worth understanding, because it explains why benchmark tables can mislead. Two models can score the same on targeted benchmarks and behave very differently on unfamiliar tasks. They plot Muse Spark 1.3 and Muse Wonder Full on in-distribution versus out-of-distribution performance, and Wonder sits above the Spark trend line: given the same in-distribution score, it does better on tasks nobody trained it for. The reason is a stronger base model. For app builders this is the difference between a model that handles your exact use case and one that handles the variation you did not anticipate, which is most of what real users do.

Safety, Stated Plainly

Meta describes Wonder as their most aligned model to date, with their strongest safeguards against CBRN and cyber misuse while preserving defensive cyber capability. It also handles private information with more care, which matters if your build touches user data. This is vendor framing and should be read as such, but the direction is right: a model that takes actions on a user's behalf needs to exercise discretion, and Meta is at least claiming to have built for that rather than bolted it on.

What It Costs on Multos

Wonder is an Elite model. At $4 input, $20 output, and $0.20 cached, the blend is $1.77 per million, a 3.54x token multiplier on Elite. That is the most expensive model in the catalog per token, and it is priced that way because it is aimed at work where the alternative is a human spending hours. The honest guidance is that you should not leave it selected by default. Use it for the hard planning turn, the architecture decision, the long debugging session, and route the mechanical work elsewhere.

Who Should Use It

If your builds are small and finish in an afternoon, Wonder is more model than you need and you will feel the cost. If you are maintaining something over months, or building an agent that has to hold a lot of context and keep working when it hits a wall, it is the strongest option here. The practical pattern on Elite is Wonder for planning and hard reasoning, Carbon for security and analysis work, and Ember or Hydrogen for the execution turns in between. That mix costs less than running Wonder for everything and produces better results than running a cheap model for all of it.

Ready to Get Started?

Muse Wonder 1.0 is live on Multos Elite. Active Memory, Active Focus, and frontier coding at $1.77 per million blended. Open the model dropdown, find it under Meta, and give it the project that has been open in a tab for a month.

Start Building Free