# What are Muse Wonder and Muse Wonder Full, Meta's new frontier models?

Date: 2026-10-13 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-meta-muse-wonder/

**Answer:** Muse Wonder and Muse Wonder Full are Meta's most capable AI models, released on October 12, 2026, and built for agents that carry out personal and professional tasks. Muse Wonder is available to developers through Meta's Model API and its Muse Code coding tool, while the larger Wonder Full is reserved for Muse Code subscribers; Meta highlights two abilities it calls Active Memory and Active Focus.

Yesterday, Monday, October 12, 2026, Meta introduced the **Muse Wonder** series: two models named **Muse Wonder** and **Muse Wonder Full**. Meta calls them its most capable models so far and says they reach the frontier, the small group of top models from the leading AI labs, on tasks done by agents. An **agent** is an AI that works toward a goal through many steps, using tools such as a browser, files or apps, rather than simply replying to one message.

The release comes from Meta Superintelligence Labs, the group Meta set up to rebuild its AI work from the ground up. It follows Muse Spark, the lab's first model family, released in April. Mark Zuckerberg called Wonder a far bigger and more advanced kind of model than Spark, according to the Latent Space newsletter's launch roundup.

## What Meta released

| | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Role | Fast frontier model | The full, larger model |
| Where | Meta Model API and Muse Code | Muse Code subscriptions only |
| API price per million tokens | $4 input, $20 output, $0.20 cached input | No API price published |

**Muse Code** is Meta's coding agent, a tool developers use to have AI write and change software. The **Meta Model API** lets developers call Meta's models from their own apps. Meta says Muse Wonder will also soon power its consumer agent product, **Muse**, starting with paying subscribers.

A **token** is a small chunk of text, often part of a word; API prices are counted per million of them. Meta's pricing page adds two details: there is no extra charge for very long prompts, and the discounted "Contributor mode," which trades lower prices for permission to train on your content, is not offered for Muse Wonder.

Meta says the models were trained at Prometheus, its gigawatt-scale data center, by scaling up both **pretraining**, the first phase where a model learns from huge amounts of text, and **reinforcement learning**, a later phase where it improves by being rewarded for good results on tasks.

## How it works: Active Memory and Active Focus

Meta presents two abilities as the heart of the release.

**Active Memory** means carrying important context about you from one conversation to the next, and pulling up the right detail when it becomes relevant. Think of an assistant who remembers that you are training for a race and avoids booking you a late dinner the night before.

**Active Focus** describes how the model keeps working toward a hard goal: trying another approach when one fails, working around obstacles and adapting to new information. Meta says it also asks questions when something is unclear and confirms with you before taking significant steps.

Meta's demonstrations show what this is meant to look like: planning a Saturday running route in a browser based on personal preferences, ordering custom merchandise designed from someone's Instagram photos, and building a wedding seating chart from a list of tricky requirements. Meta also stresses expertise in law, finance and health, and strong coding performance on long software projects.

## How good is it?

Meta's own charts claim top results on agent tasks, on memory tests and on coding. Those are the company's measurements. An early independent view comes from Artificial Analysis, as reported by Latent Space: on its Intelligence Index, a combined score across ten tests, Wonder Full scores 56 and Wonder 52, up from 48 for Muse Spark 1.3. Anthropic's Claude Opus 5.5 leads that index at 58. In other words, outside testers place Meta near the top for the first time, though not in first place.

The launch was not seamless. Latent Space reports a bug that used up Muse Code subscribers' usage limits far too quickly on launch day; Meta fixed it the same day and reset everyone's quotas.

On safety, Meta says the Wonder family was trained to resist adversarial instructions and to handle private information carefully, which matters for an agent that may read your messages, calendar and documents. It calls Muse Wonder its most aligned model and says it has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse. A fuller safety report is promised but not yet published.

## A worked example

A personal agent with memory works best when you state your preferences once and let it apply them. A first request might look like this:

```prompt
Remember these preferences for future planning: I run in the morning,
I avoid hills, and I prefer routes of 8 to 10 km near water.
Now plan my run for this Saturday near my home, check the weather,
and ask me before adding anything to my calendar.
```

The first sentence gives Active Memory something worth keeping. The last line uses Active Focus's habit of checking in before an action that changes your accounts.

> **Try it:** If you have Muse Code or API access, start one conversation by stating three preferences, then open a new conversation a day later and ask a question where those preferences should matter. Note whether the model recalls them without being reminded.

## Why it matters

Muse Wonder is Meta's first model family that it, and early outside testers, place close to the best from OpenAI and Anthropic, with a price, $4 and $20 per million tokens, aimed at the same paying developers. Meta's focus on memory and persistence also shows where the competition is heading: less toward clever single answers, more toward agents trusted to work on someone's behalf over days and weeks.

## How can you try Muse Wonder, and what should you check first?

Developers can call Muse Wonder through the Meta Model API today, under the model name `muse-wonder-1.0`, or use it inside Muse Code. Wonder Full requires a Muse Code subscription. Everyday users will need to wait for Muse, where Meta says Wonder is coming soon, subscribers first.

Before relying on it, check three things. First, test it on your own tasks rather than trusting launch charts. Second, look at what it remembers about you and how to view or delete that memory, since persistent memory is the feature most likely to surprise you. Third, keep approvals on for actions such as purchases or sending messages until you have seen how it behaves.

## Sources

- Muse Wonder: A Frontier Model For Personal & Professional AI Agents: https://research.meta.ai/blog/muse-wonder
- Pricing and rate limits - Meta Model API: https://dev.meta.ai/docs/pricing-rate-limits
- Meta Muse Wonder & Wonder Full: MSL's Claude 5.5-tier Frontier Model: https://latent.space/p/ainews-meta-muse-wonder-and-wonder