Meta Releases Its First Frontier Model, Muse Wonder, and Asks the Public to Please Test It Themselves
After fourteen months of restructuring, Meta ships a model that scores 'comparable' to the leaders, with the benchmark tables mostly left as an exercise for the reader.
On Monday, Meta’s Superintelligence Labs released Muse Wonder and Muse Wonder Full, which TechCrunch describes as the company’s first attempt at a frontier model. This is, by the calendar, a notable moment. Alexandr Wang restructured Meta’s AI efforts in August 2025. In the fourteen months since, Meta has shipped a lightweight model called Muse Spark and an agent called Muse, which is a perfectly respectable amount of output for an organization that has also been busy being restructured.
Now, the frontier has been reached. Or approached. Or, per Meta’s own model card, matched “comparably.”
Some precise facts, because the facts here are doing a lot of work. Per TechCrunch, Meta says both models score comparably to Claude Fable 5.1 and GPT-6 Astra, and that Muse Wonder beats Fable, Opus and Astra on MemoryBench, a benchmark for continual learning and memory construction. The article reports no specific numerical scores. The launch materials, according to AI News Weekly, don’t publish the per-benchmark numbers for the memory claim either.
Independent numbers do exist. AI News Weekly, summarizing Wired’s reporting, puts Muse Wonder Full at 56 and Muse Wonder at 52 on the Artificial Analysis Intelligence Index, against 58 for Claude Opus 5.5. That places Meta’s flagship just below the best model on the list and level with a Sonnet. It is the AI equivalent of finishing the marathon in the top group, with a good time, at a different race than you said you were running.
The Next Web adds that, per Wang, the model matches ChatGPT and Claude on coding and on financial and legal work, a claim delivered without published benchmarks. Wang has urged people to test the model themselves. This is a bold new product strategy: ship the model, publish the vibes, and outsource the evaluation section to the customer. Some companies call that a launch. Others call it open-book grading, except the book is yours.
The feature Meta is leaning on hardest is memory. The company calls it “active memory,” the ability to retain and synthesize information from users over days, weeks, months “and beyond.” Research scientist Kavya Kopparapu said Meta put “a tremendous amount of effort into making sure your model can remember.” There is also “active focus,” described by AI News Weekly as a feature meant to keep multi-step agents from drifting between apps.
It is worth sitting with the pitch for a moment. This is Meta, a company whose entire commercial history is a relationship built on remembering things about you, announcing a model whose headline feature is remembering things about you, “and beyond.” The beyond is unspecified. We assume it’s a roadmap item.
The pricing is where Meta decided to be legible. Muse Wonder is on the Meta Model API at $4 per million input tokens and $20 per million output tokens, which TechCrunch says is roughly in line with GPT-6 Sol or Opus 5.5. The larger Wonder Full is not served through the API at all. It is available to Muse Code subscribers, with tiers listed by AI News Weekly at $50 a month and $200 a month.
So the cheaper model is for sale to developers, and the stronger model is for sale to people who also bought the coding product. It is a very Meta structure: the best thing is behind the thing you already pay for, and the thing you already pay for is behind the other thing.
Both models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year. The data center is one of those facilities that gets named in press releases and then spoken of by everyone in the same hushed voice normally reserved for aircraft carriers. Somewhere in Ohio, a very large building has learned to remember things, and has been asked to please not mention that to the corn.
Now the open question, which is quite literally about openness. Meta has not said whether the Wonder weights will be published. The Next Web notes this matters under the EU AI Act: a genuinely open model gets an exemption from some documentation duties, but the license has to be real, and models with systemic risk face full obligations regardless of licence. Meta, per the article, has declined to sign the code of practice designed to implement those obligations.
For now, the question of whether Wonder is open has the same answer as the question of when it will be: Meta hasn’t said.
What we can say for certain is this. Meta has a frontier-class model. It sits slightly below the leader on the one independent index we could find. It remembers things. It costs the same as its competitors. It was trained in Ohio. And its maker would like you to evaluate it yourself, which, to be fair, is exactly what an industry full of vendor-supplied benchmarks has spent three years telling everyone not to do.
Wang may be right. Perhaps the only benchmark that matters is the one you run on your own Tuesday. But for the record, the model will remember how that went.
Sources: