Muse Wonder

Learning Objectives

• Explain what Muse Wonder is and how it differs from Muse Spark
• Understand what Meta means by Active Memory and Active Focus
• Judge Meta's benchmark claims, and know where the two Wonder models can and cannot be used

What Is Muse Wonder?

Muse Wonder is Meta's first model aimed at the top of the field. Meta Superintelligence Labs released it on October 12, 2026, together with a larger sibling, Muse Wonder Full. Meta calls them its most capable models and "the next step on our scaling ladder."

They sit above Muse Spark, which Meta designed to be light and quick to respond. The Wonder models are larger and slower to run, and they are aimed at hard, long-running work: coding projects, research, and personal agents that act across a user's apps over days. Meta says both were trained at its Prometheus data center in central Ohio, a site of about a gigawatt that came online in 2026. The release is the first frontier model since Alexandr Wang rebuilt Meta's AI effort in August 2025.

Meta has not announced open weights for either model. Like Muse Spark, they are served only from Meta's cloud.

Active Memory and Active Focus

Meta frames the release around two abilities a personal agent needs.

• Active Memory is carrying what matters about a user from one conversation to the next, and pulling the right detail back up when it becomes relevant. Meta points to the MemoryBench benchmark, where it says Muse Wonder scores above Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra.
• Active Focus is staying on a hard task: trying another approach when one fails, working around blockers, and asking for clarification or checking in before an important action. Meta measures this with its own FocusBench and with SWE-Marathon, which tests long software projects.

Meta also argues the Wonder models do better than their benchmark scores suggest on tasks they were not trained for. Its evidence is a chart comparing Muse Wonder Full with Muse Spark 1.3 on targeted and untargeted tasks, which it attributes to a larger, stronger base model.

Muse Wonder and Muse Wonder Full

The two models differ mainly in where you can use them.

• Muse Wonder is available through the Meta Model API and inside Muse Code, Meta's terminal coding agent. Meta says it will also power the Muse personal agent soon, starting with subscribers, and will be used to build smaller specialized models for everyone else.
• Muse Wonder Full is the larger model. It is available only to Muse Code subscribers and is not offered through the API.

Meta says the family has its strongest safeguards yet against chemical, biological, radiological and nuclear misuse and against cyberattacks, while keeping advanced cyber abilities available for defensive work under its Meta Superintelligence Scaling Framework. A full safety report has been promised but was not yet published as of October 13, 2026.

Pricing

That is the same price as Claude Opus 5.5 and much less than GPT-6 Astra and Claude Fable 5.1, which both charge 10 dollars in and 50 dollars out. It is about three times Muse Spark's standard input price and nearly five times its output price, so Muse Spark remains the cheaper choice for high-volume, simpler work.

Strengths

• Built for memory over time, which is the weak point of most assistants once a conversation ends
• Priced like Opus 5.5, well below the premium tier of OpenAI and Anthropic models
• Designed for long tasks, with training on long-running coding and agent work rather than single answers
• One family across Meta's products, from the API to Muse Code to the Muse agent

Limitations and Considerations

• The evidence is all vendor-reported. Meta picked the benchmarks and wrote some of them. Test it on your own tasks before trusting the headline comparisons
• The larger model is locked inside Muse Code. Developers building their own apps can only reach Muse Wonder, not Muse Wonder Full
• No safety report yet. Meta has described its safeguards but not yet published the full evaluation it promised
• Not yet in the Muse agent. Meta says that is coming, starting with subscribers, but gave no date
• Closed and cloud-only. There are no downloadable weights, unlike Meta's Llama models and Muse Glimmer
• Memory is a privacy question as well as a feature. A model built to remember you for months holds more about you; check what is stored and how to delete it before connecting personal accounts

Best Use Cases

Getting Started

1. Get access to the Meta Model API and select Muse Wonder
2. Pick a task you already run on another frontier model, ideally one that spans several sessions
3. Run the same inputs through both models and compare quality, cost and how much context each one keeps
4. To try Muse Wonder Full, use it inside Muse Code on a real repository
5. Keep a person reviewing actions that touch money, accounts or other people until you trust its judgment

Key Takeaways

• Muse Wonder, released October 12, 2026, is Meta's first frontier model, with a larger Muse Wonder Full alongside it
• It is built around Active Memory and Active Focus: remembering a user over weeks and staying with long tasks
• The API price is $4 per million input tokens and $20 per million output, the same as Claude Opus 5.5
• Muse Wonder Full is only available inside Muse Code, not through the API
• Every benchmark result so far is Meta's own, and the promised safety report had not been published as of October 13, 2026
• Muse Spark remains Meta's faster, cheaper model and still powers Meta AI and the Muse agent for now