Meta's First Frontier Model Is Closed, and Its Best Version Is Not on the API

Meta released its first frontier model on 12 October, and the detail that matters is not the benchmark claim. Muse Wonder and Muse Wonder Full are closed. There are no weights to download, no licence to read, no fallback if the terms change. The company whose entire AI identity was built on shipping weights has, at the top of its line, stopped shipping them.

Both models take text, image, video and audio in and return text, with a 1M-token context window. Meta trained them at Prometheus, its gigawatt-scale data centre in central Ohio, which came online earlier this year. How you reach them differs sharply:

The pricing says Meta thinks it has caught up

$4 in and $20 out puts Muse Wonder at roughly the same cost as GPT-6 Sol and Claude Opus 5.5. That is a positioning statement. A lab that believed it was a tier behind would undercut; Meta has priced at parity and asked buyers to accept its claim of parity.

The claim itself is substantial. Meta says the Wonder models reach state-of-the-art results on agentic, memory and coding benchmarks, and its comparison table puts them against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 across the Personal Superintelligence Index, Harvey's legal agent benchmark, Vals Finance Agent V2, HealthBench Pro, OSWorld V2.1, Terminal-Bench 4.0, SWE-Marathon, DeepSWE 1.1, ProgramBench and the Artificial Analysis index. On MemoryBench, which tests continual learning rather than one-shot answering, Meta says Muse Wonder beats all three.

Two features carry the pitch. Active Memory keeps relevant context across separate conversations. Active Focus is meant to keep a model working a goal through failure — recovering from errors, asking for clarification instead of guessing. Research scientist Kavya Kopparapu said in the launch video that the team put

a tremendous amount of effort into making sure your model can remember what's important to you

over days, weeks and months. That is the right thing to be building for agents, and it is the one area where Meta's distribution — an assistant already inside products a billion people open daily — is a genuine advantage over a lab selling API tokens.

Every number is Meta's own

Nothing here has been independently evaluated. The scores come from Meta's model card and its developer pages; no third-party benchmark of Muse Wonder existed at launch, and the promised full safety report is listed as arriving "soon." Forkast reported that Meta's own research blog index does not list the Muse Wonder post at all, with Muse Spark 1.1 from 9 July still the most recent entry, and that the launch reached most people through a Threads post from Mark Zuckerberg and secondary coverage.

The model card does publish risk-assessment figures, and they are not shy ones. Muse Wonder Full is reported at 88.4% on CyberGym and 58.9% on ExploitBench — offensive-security capability scores high enough that the deferred safety report is the document buyers should actually want. The card also reports 87.3% on MASK, which tests honesty under pressure, a 0% harmful-action rate on agentic misalignment, and refusal rates of 98.2–98.7% on BioTIER's English, multilingual and adversarial splits.

So the capability numbers and the safety numbers are both the vendor's, and the independent check on either does not exist yet. For a model being sold into law, finance and health — the three domains Meta named — that is the live problem. An enterprise buyer evaluating Muse Wonder this week has Meta's word and a price list.

Open and closed, running at once

Meta has not abandoned open weights entirely. Muse Glimmer 30B shipped on 9 August under an MIT-style licence, and the small end of the line stays open. What has changed is that the frontier no longer is. Llama's argument was that openness was the strategy, not a concession — that Meta would win by commoditising what rivals sold. Muse Wonder Full, locked behind a $200 subscription and absent from the API, is the opposite argument: the best thing we have is worth gating, and we will gate it the way everyone else does.

This is the first frontier release since Alexandr Wang reorganised Meta's AI effort in August 2025, and on the evidence of the launch, that reorganisation produced a product company. Meta is now competing on the terms OpenAI and Anthropic set, with the same business model and the same opacity. Whether it has the model to match is a question its own numbers cannot settle.

Sources: Meta Newsroom: Introducing Muse Wonder and Muse Wonder Full · Meta AI Research: Muse Wonder · Muse Wonder model card · Meta developer docs: Muse Wonder · TechCrunch: Meta releases its first frontier model, Muse Wonder · Forkast: Meta Ships Muse Wonder, Its First Frontier Model — and Leaves the Research Blog Empty