Meta has spent more than a year restructuring its AI effort, and the first thing it has to show at the frontier is a model that promises to remember your birthday. On Monday, Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, the company's first attempt at a frontier model.
The pair sit above Muse Spark, the lightweight, low-latency model Meta released earlier. Wonder is larger and, by Meta's account, more capable, with higher benchmark scores and better results on complex work such as coding.
What Meta is claiming
The headline claim comes from Meta's own testing, presented in the model card. By those numbers, Wonder and Wonder Full reach benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
That is a company grading its own homework, which is how nearly every launch works. Nothing in the announcement points to independent verification, so treat the comparison as Meta's position and not a settled result.
Meta's blog post frames the release in suitably grand terms: "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence." The phrase "scaling ladder" suggests there are more rungs planned, though the post as reported does not say what they are.
Pricing and access
The two models are not offered the same way.
| Muse Wonder | Muse Wonder Full | |
|---|---|---|
| Meta Model API | Yes | No |
| Muse Code subscribers | Not specified | Yes, exclusively |
| Powers the Muse agent | Yes | Yes |
Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. That puts it roughly level with GPT-6 Sol and Opus 5.5. Meta is not undercutting anyone here, which fits a model positioned as a peer and not a bargain.
Wonder Full is the more exclusive of the two. Only Muse Code subscribers can use it, and it will not be served through the API. Developers who want to build on the larger model through an API call will have to make do with Wonder.
The pitch: active memory
If there is one idea Meta wants attached to these models, it is memory. The company is interested in personal assistants that can carry out tasks over a long period, and it put particular weight on the models' ability to retain and synthesize what a user tells them. Meta calls this "active memory."
To back it up, Meta pointed to MemoryBench, a benchmark that challenges continual learning and memory construction frameworks. In the company's testing, Muse Wonder scores higher on MemoryBench than Fable, Opus or Astra.
Research scientist Kavya Kopparapu put the ambition plainly in a video accompanying the launch: "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond."
The claim is easy to say and hard to check. A model that scores well on a memory benchmark is not necessarily one that remembers the right things about you after three months. That is the question users will answer once the Muse agent is running on these models.
Why this release matters for Meta
The timing matters more than the specs. These are the first frontier models to come out since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year. In the time since, Meta shipped the lightweight Muse Spark and the Muse agent, but it held back from the kind of state-of-the-art releases that OpenAI and Anthropic are known for.
Wonder ends that pattern. Whether it ends it convincingly depends on how the benchmark claims hold up outside Meta's own model card.
There is also a hardware footnote. According to Meta's post, the models were trained at the Prometheus data center in central Ohio, which came online earlier this year. So the facility has produced a flagship model within months of opening, which is a tidy story for a company that has been building out compute.
What we do not know
The announcement leaves several things unsaid, at least in what has been reported:
- Independent results. The benchmark comparisons come from Meta's testing.
- Differences between the models. Wonder Full is described only by its restricted access, and the reported material does not spell out how it differs from Wonder in capability.
- Real-world memory. MemoryBench is a measure, not a lived experience.
What is clear is the strategy. Meta is pricing Wonder alongside the best-known rivals, keeping its largest model behind a coding subscription, and betting that an assistant that remembers you is the feature that separates it from a field where benchmark scores are increasingly bunched together. It is a reasonable bet, assuming the assistant remembers the right things.