Meta launches Muse Wonder: a frontier model that is no longer open
Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Scores of 52 and 56 on the Artificial Analysis Index, but the weights can no longer be downloaded.
On October 12, 2026, Meta released two new models: Muse Wonder and Muse Wonder Full. These are the first frontier-class models from Meta Superintelligence Labs (MSL), the AI research organisation led by former Scale AI CEO Alexandr Wang.
Neither is open-weight. That is a big shift for Meta.
1. Two models, two very different access paths
Meta shipped two versions with sharply different availability:
| Muse Wonder | Muse Wonder Full | |
|---|---|---|
| Access | Meta Model API, Muse Code, OpenRouter | Muse Code only |
| API price | $4 input / $20 output per 1M tokens | No public API price |
| Subscription | None | $50/month (Power) or $200/month (Maximum) |
Muse Wonder is the one you can build with. Call it through Meta's API using model ID muse-wonder-1.0, or through Muse Code and OpenRouter.
Muse Wonder Full, the most capable version, is reserved for Muse Code subscribers. No API access, no pay-as-you-go pricing.
This is an unusual business decision. AI companies normally sell their best model through an API. Meta did the opposite.
2. Performance: close to the top
On the Artificial Analysis Intelligence Index, which synthesises nine evaluations into a single score:
- Muse Wonder: 52
- Muse Wonder Full: 56
- Claude Opus 5.5: 58
- GPT-6 Astra: around 56
Muse Wonder Full sits two points behind Opus 5.5. Fourteen months ago, Meta was a mid-tier lab with open models. Today, on paper, it is knocking on the door of the best model in the world.
Important caveat: Meta has not disclosed a parameter count. These scores come from third-party evaluation, not Meta's own claims.
3. What it can do
Both models:
- Accept text, images, audio and video as input
- Produce text output
- Support a context window of approximately 1 million tokens
- Rate limits: 3,000 requests and 4 million tokens per minute
Meta says both were developed using its Prometheus AI data centre, combining large-scale initial training with reinforcement learning to improve their ability to complete complicated tasks.
The headline capability Meta emphasises: Active Memory and Active Focus: a model that keeps track of what matters to you and works toward your goals at work and in life.
4. The big issue: no longer open-weight
This is the part that matters most.
Meta is famous for open models. Llama reshaped the AI landscape because anyone could download and run it. Muse Glimmer, their 30B model, is still open under Apache 2.0 and optimised to run on personal hardware.
But Muse Wonder and Muse Wonder Full are not open-weight. The weights cannot be downloaded. You cannot self-host. You cannot modify them.
Meta now runs two separate tracks:
- Hosted frontier models (Muse Wonder, Wonder Full), access through paid services
- Open local models (Muse Glimmer 30B), download and run yourself
Wang said Meta remains committed to open source. But this launch shows where they are putting their best models.
5. Data terms you need to read
Meta offers two data-use arrangements:
- Standard mode: content you submit is not used to train its models. The pay-as-you-go API is available only in this mode.
- Contributor mode: you may share data in exchange for better terms.
For enterprise developers, this means that if you hold sensitive data, you need to understand exactly which mode you are in and what happens to your data.
Meta's Model API Terms of Service permit commercial application development and give customers ownership of their submitted inputs and generated outputs. But access to Wonder itself remains governed by Meta's hosted service rather than an open-weight licence.
What it means for you
This launch changes two things: your options, and your assumptions about Meta.
If you are a developer or product owner:
- Meta is now a real option. If you build AI agents, Muse Wonder at $4/$20 is competitive pricing against Claude Opus 5.5 and Gemini 4 Argon. Worth testing.
- But do not assume it is open. If you need a model you can self-host (for privacy, long-term cost, or control), Meta does not offer that at the frontier. Muse Glimmer 30B still exists, but it is not the same class.
- Read the data terms before sending anything. Standard vs Contributor mode is a difference that can affect your compliance posture.
- Do not pay for Wonder Full unless you actually use Muse Code. $200/month only makes sense if your coding work is heavy enough.
For everyday users: Muse Wonder will power features in Meta's Muse app, initially for paid subscribers. If you use Muse, you will feel the difference without doing anything.
Want to know which AI tools suit your work right now? See my picks on the AI Tools page.
Sources
- Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance (VentureBeat)
- Meta unveils new frontier-class AI models to compete with Anthropic, OpenAI (Axios)
- Muse Wonder: A Frontier Model For Personal & Professional AI Agents (Meta AI Research)
- Meta's Most Advanced AI Model Isn't Open-Weight (TechGenyz)
- Meta's Muse Wonder Scores Big and Delivers Little (DEV Community)