# Muse Wonder and Muse Wonder Full: Meta unveils its first frontier AI models

> On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, its most capable models for AI agents, coding and document work. Muse Wonder is available through the Meta Model API and Muse Code; Wonder Full is for Muse Code subscribers only. Wonder Full scored 56 on the Artificial Analysis index, second only to Claude Opus 5.5. The model is coming to AI2Face soon.

Published 2026-10-13 · Updated 2026-10-13
URL: https://ai2face.com/news/muse-wonder

On October 12, 2026, Meta introduced **Muse Wonder** and **Muse Wonder Full**, which the company calls its most capable models and “the next step on our scaling ladder.” They build on the lightweight Muse Spark line, were trained at Meta's gigawatt-scale Prometheus data center, and target agentic work: personal tasks, coding, and legal, financial and health documents.

## What Muse Wonder can do

Meta highlights two new capabilities:

- **Active Memory** — the model remembers what matters about you and pulls the right details into later conversations. On MemoryBench, Meta's internal memory test (64K–100K-token conversations with conflicting and updated facts), Muse Wonder beats Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra, according to Meta.
- **Active Focus** — the model sticks with a hard task on its own: it tries alternative approaches, works through blockers, adapts to new information and checks in with you before important actions.

The model works in the browser and across other apps. In Meta's demos it builds a 3D game from photos of a kitchen, puts together a running plan from personal preferences, orders custom merch from Instagram photos and arranges wedding seating. For coding, it edits multiple files, runs tests and iterates on projects ranging from small apps to complex systems.

According to the model card, Muse Wonder takes text, images, video and audio as input and outputs text only. The context window is 1 million tokens. It can't generate images: Meta has a separate model, [Muse Image](/models/muse-image), for that.

## Muse Wonder on the Artificial Analysis leaderboard

On the composite Artificial Analysis Intelligence Index v4.3 (independent runs by Artificial Analysis, including Terminal-Bench 4.0, SciCode and Humanity's Last Exam), Muse Wonder Full ranks second and Muse Wonder ties for sixth:

| Model | Developer | AA Index score |
|---|---|---|
| Claude Opus 5.5 | Anthropic | 58 |
| **Muse Wonder Full** | Meta | **56** |
| Claude Fable 5.1 | Anthropic | 53 |
| GPT-6 Astra | OpenAI | 53 |
| Gemini 4 Argon | Google | 53 |
| **Muse Wonder** | Meta | **52** |
| GPT-6.1 Sol | OpenAI | 52 |
| Claude Fable 5 | Anthropic | 50 |
| Muse Spark 1.3 | Meta | 48 |

The family's progress shows release by release: Muse Spark scored 31, Spark 1.1 — 34, Spark 1.2 — 40, Spark 1.3 — 48, Wonder — 52 and Wonder Full — 56. Meta compares models at maximum reasoning effort. Some benchmarks in the model card are Meta's internal tests (FocusBench, MemoryBench, PSI), and nobody has independently verified them yet.

## Where Muse Wonder and Muse Wonder Full are available

- **Muse Wonder** — on the Meta Model API (ID `muse-wonder-1.0`, compatible with OpenAI clients) and in Muse Code, Meta's coding agent. API pricing: $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens — the same as Claude Opus 5.5.
- **Muse Wonder Full** — Muse Code subscriptions only, on the Power Usage and Maximum Usage plans. Maximum Usage gives 80x the limit of the base Everyday Usage plan. Wonder Full isn't sold through the API.
- **Muse app** — Muse Wonder will soon power the app, rolling out to paid subscribers first.

Meta isn't releasing the model weights. On safety, the company calls Muse Wonder its “most aligned” model yet: in the model card, Wonder Full refuses 98.7% of high-risk biology prompts and takes harmful actions 0% of the time on Agentic Misalignment.

## Muse Wonder in AI2Face

Muse Wonder is coming to AI2Face soon, together with text generation: the service is adding the top language models, including Meta's new release. You'll be able to write texts, prompts and descriptions in the same place where you create images and videos. The price is shown before each generation.

Muse Wonder itself only outputs text. Images and video in AI2Face come from specialized models: [Nano Banana Pro](/models/nano-banana-pro), [GPT Image](/models/gpt-image) and [Veo 3.1](/models/veo-3-1). For a comparison of image models, see our overview of [AI image generators](/blog/nejroseti-dlya-izobrazhenij).

## FAQ

### What's the difference between Muse Wonder and Muse Wonder Full?

Full is the most capable variant: 56 on the Artificial Analysis index versus 52 for Muse Wonder. Muse Wonder is sold through the API, while Wonder Full is only available with a Muse Code subscription, starting with the Power Usage plan.

### How much does Muse Wonder cost?

On the Meta Model API it's $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. That's the same as Claude Opus 5.5.

### Can Muse Wonder generate images?

No. The model understands text, images, video and audio, but only outputs text. Meta has a separate model, Muse Image, for images.

### Can I download Muse Wonder?

No. Meta hasn't announced open weights: the model runs through the Meta Model API, Muse Code and, soon, the Muse app. Muse Wonder is also coming to AI2Face soon.

## Sources

- https://research.meta.ai/blog/muse-wonder
- https://research.meta.ai/static/muse-wonder-model-card
- https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- https://dev.meta.ai/resources/blog/build-with-muse-wonder