Meta says Muse Wonder reached the AI frontier. Its own scorecard says Claude still edges it out.

Meta's first frontier model, fourteen months in the making

On 12 October 2026, Meta's Superintelligence Labs released Muse Wonder and Muse Wonder Full, the first models Meta has called a genuine attempt at the AI frontier since Alexandr Wang restructured the company's AI effort fourteen months earlier, in August 2025. Until now Meta had shipped the lighter Muse Spark model and the Muse personal-agent app, but had avoided the kind of state-of-the-art release associated with OpenAI, Anthropic, and Google. Both new models were trained at Meta's Prometheus data center in central Ohio, which came online earlier in 2026, and both read text, images, video, and audio, outputting text with a context window of 1 million tokens at inference, according to Meta's own model card. "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," Meta's launch post says.

What the model card actually shows

Meta backed the release with a model card comparing Muse Wonder and Muse Wonder Full against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5 across 14 benchmarks, each run at maximum reasoning effort. The real numbers tell a more complicated story than "frontier." Muse Wonder Full leads clearly on memory, scoring 85.5 on MemoryBench, Meta's own internal test of recalling facts buried in 64,000 to 100,000 token conversations, ahead of Claude Opus 5.5's 81.5 and GPT-6 Astra's 81.4. It also tops the Personal Superintelligence Index, an internal composite of seven agent and knowledge-work tests, at 59.1 against Opus 5.5's 55.2. On the long-horizon coding benchmarks DeepSWE 1.1 and SWE-Marathon, Wonder Full's scores of 79.0 and 57.5 both beat every rival model Meta tested.

But Wonder Full does not win everywhere, and Meta's own table says so. On HealthBench Pro, a set of 525 professional-grade health questions graded against physician-written rubrics, Claude Opus 5.5 scores 65.6 and GPT-6 Astra 65.3, both ahead of Wonder Full's 62.3. On OSWorld V2.1, a long-horizon computer-use benchmark where an agent operates a real desktop through screenshots and clicks, Opus 5.5 scores 81.8 against Wonder Full's 78.4. On ProgramBench, which grades an agent's ability to reconstruct a full program from scratch in a clean room, Opus 5.5 posts 70.5 to Wonder Full's 59.0, the widest gap on the entire table.

The one score Meta did not grade itself

The most telling row sits near the bottom of the card: the Artificial Analysis Intelligence Index, a composite benchmark run and scored by Artificial Analysis, an outside evaluation firm, not by Meta. Every other number on the card comes from Meta's own internal test harness, even when it borrows a third party's questions. On this one, where the scoring itself is independent, Claude Opus 5.5 posts 58 against Muse Wonder Full's 56, with GPT-6 Astra and Claude Fable 5.1 tied at 53. Meta's model card is upfront that it ran most of these comparisons itself, noting that for rival models "prompts, tools, and runtime may not be optimally tuned." That caveat cuts both ways. It is an honest disclosure, and a reminder that nearly every other number on the table is Meta grading its own homework.

Real model, locked behind a subscription

None of this is free to try. Muse Wonder costs 4 dollars per million input tokens and 20 dollars per million output tokens through the Meta Model API, in the same range as GPT-6 Sol and Claude Opus 5.5. Muse Wonder Full, the stronger of the two models and the one that leads the benchmark table, is not sold through the API at all. It is available only to subscribers of Muse Code, Meta's coding product, or eventually to paying users of the consumer Muse app, which costs 20 or 100 dollars a month and previously ran on the smaller Muse Spark model. Meta says Wonder will "soon power experiences in Muse," rolling out first to paying subscribers, with no date given. The free tier of Muse gets neither model yet. Meta has also promised a separate safety report "available soon" covering further evaluations beyond the model card's own chemical, biological, and cybersecurity risk tables, which it says keep the model at a "moderate or lower" risk level once mitigations are applied.

That subscription wall sits inside a personal AI agent Meta has spent over a year building, most recently with the keychain-sized Muse Charm hardware Zuckerberg showed off at Connect. Wonder is the model meant to make that hardware worth wearing: Active Memory is what lets Muse recall a user's preferences across sessions, and Active Focus is what lets it keep working on a task instead of stalling at the first obstacle. Research scientist Kavya Kopparapu put it 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."

Our take

Muse Wonder is Meta's first real swing at the frontier, and on its own evidence it is not a bluff. The memory and long-horizon coding numbers are genuinely ahead of Claude and GPT-6 on the tests Meta chose to run. What it is not, by Meta's own table, is simply the best model on the market. Claude Opus 5.5 beats it on health questions, computer use, program reconstruction, and the one composite score an outside firm scored instead of Meta. Meta's framing, a major leap toward personal superintelligence, is the kind of line this desk's hype check exists for. Credit the real gains in memory and coding, and treat competitive with the best models out there as the more honest description of where Wonder actually lands.