Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, calling them its most capable models yet. The pitch is less about raw intelligence and more about an assistant that keeps what it learns about you.
That matters because Muse Wonder is built to remember your context between conversations and keep pushing on hard tasks. If it works, you stop re-explaining yourself every time you open a chat.
TL;DR: Muse Wonder is Meta’s new top model, built around remembering you across conversations and finishing long tasks. Developers can use it today through the Meta Model API, while everyday users have to wait for an unspecified rollout in the Muse app, subscribers first. Meta claims leading results but publishes few numbers for them.
What Muse Wonder actually is
Muse Wonder is the next step after Muse Spark in Meta’s model lineup, according to Meta’s announcement. Muse Wonder Full is the stronger variant, and it is limited to Muse Code subscribers for now.
The model card adds the technical basics. Both models take text, image, video and audio as input, but they only answer in text. Context length is 1M tokens at inference.
| Spec | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Release date | October 12, 2026 | October 12, 2026 |
| Available now | Meta Model API, Muse Code | Muse Code subscribers only |
| Input | Text, image, video, audio | Text, image, video, audio |
| Output | Text only | Text only |
| Context length | 1M tokens | 1M tokens |
A text only output is a small limit that is easy to miss. Whatever the Muse app eventually does with Wonder, the model itself will not speak or draw.
Active Memory and Active Focus are the real pitch
Meta describes Active Memory as carrying important context about you across conversations and pulling up the details that matter for the task at hand. Active Focus is the other half: the model keeps working toward a hard goal, tries different approaches when it hits a blocker, and checks in before important actions.
Think of asking an assistant to plan a Saturday run. Today you repeat your pace, your route habits and your schedule each time. Meta’s own demo does this kind of planning across a browser and other apps from your stated preferences.
The catch is that memory only helps if you trust what it kept. The announcement does not describe memory limits or how you would review and clear what the model stores, so that question stays open.
How strong are the Muse Wonder benchmarks?
Meta says Muse Wonder reaches state of the art on personal and professional agent tasks, memory and coding. The model card compares it with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all set to maximum reasoning effort.
What it does not do is print scores for most of those tests. The charts are on Meta’s page, and the card lists the benchmark names without numbers for the general results. Until independent testers repeat them, treat the leading claim as Meta’s own.
The numbers that are published come from the cyber and safety sections. Muse Wonder Full scored 88.4% on CyberGym and 58.9% on ExploitBench. On ExploitGym it captured 92 flags, but only 21 used the intended vulnerability, and Meta says the 21 is the figure comparable to the published ranking.
Who can use Muse Wonder today
Developers get the first access. Muse Wonder is live in the Meta Model API and in Muse Code, and Wonder Full needs a Muse Code subscription.
Everyone else waits. Meta says Muse Wonder will soon power experiences in Muse, rolling out to subscribers first, and gives no date. It also plans to use the model to build more specialized models for Muse users, and for now it does not mention the Meta AI app by name.
For a look at how another lab prices a smaller model, see our Claude Haiku 5.5 pricing breakdown.
Safety limits worth reading before you rely on it
Meta says Muse Wonder has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse. It also calls the model its most aligned to date, with attention to how agents handle private information such as messages, calendars and documents.
The model card is more cautious. It lists weak spots in 3D and spatial reasoning and warns that answers can be inaccurate. It says the model is not intended for people under 18, a point that echoes the concerns in our report on ChatGPT parent alerts. A fuller safety report is promised soon, with no date.
What to watch before the Muse app gets it
The useful test is not a benchmark chart. It is whether Active Memory feels helpful in the Muse app once it arrives, and whether Meta lets you see and clear what it stored.
Not a dramatic change, but a telling one: Meta is competing on how well an assistant knows you, not just how well it answers. For now the story is developer access, a lot of claims, and a consumer rollout still waiting for a date.
