Meta's New Muse Wonder Full Beats Claude Fable 5.1 on Independent Testing. It's Not the Model Developers Can Actually Call.

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12 -- the pair it is calling its first genuinely frontier-class models, after six months of Muse Spark releases that Meta's own AI chief, Alexandr Wang, had already conceded weren't "at the tier of the leading frontier models." Both are pitched at agentic work, personal and professional, with Meta framing the release as the moment it stopped rebuilding and started pushing the frontier.

Access to the two models is not symmetric. Muse Wonder is available now through the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens, and also powers Muse Code and (soon) the Muse consumer app. Muse Wonder Full is restricted to Muse Code subscribers only -- it is not served through the API at all, and Meta's own announcement lists no price for it. Wang told reporters that whatever Meta eventually charges for broader access will stay "meaningfully below" what OpenAI and Anthropic charge for comparable capability -- a promise, not yet a number.

That access split matters because the two models are not close to each other in independent measurement. Artificial Analysis, which runs the only third-party Intelligence Index this newsroom treats as a score rather than a claim, puts Muse Wonder (max) at 52, ranked 9th of 219 comparable configurations -- essentially tied with Claude Fable 5.1 and GPT-6 Astra, both already at 53 on this site's own board. Muse Wonder Full (max) scores 56, 4th of 699 -- clear of Fable 5.1 and Astra, and trailing only Claude Opus 5.5's 58. Meta's own model card describes performance "comparable to Fable 5.1 and GPT-6 Astra" -- which is an accurate description of Wonder, the model anyone can actually call through the API, and an understatement of Wonder Full, the model almost nobody can.

The headline feature Meta is actually selling, though, is memory. Active Memory is built to retain and synthesize what a user tells it over weeks or months rather than one conversation; Active Focus is meant to keep the model working toward a goal through a hard multi-step task, checking in with the user before it acts on anything that matters. Research scientist Kavya Kopparapu put it plainly in Meta's launch video:

“We put a tremendous amount of effort into making sure your model can remember what's important to you.” — Kavya Kopparapu, Meta research scientist, launch video

Meta backs that with a specific claim: on its own internal MemoryBench evaluation, Muse Wonder scores higher than Fable 5.1, Claude Opus, and GPT-6 Astra. No outside lab has reproduced that result, and MemoryBench is Meta's benchmark, run on Meta's models -- the same structure as every vendor-run benchmark this newsroom declines to treat as a score.

That September episode is worth naming rather than re-litigating here: Meta's Muse personal-agent product has already drawn a password-change bug and competitive-pressure questions over its launch timing, and the dispute over unauthorized message access pushed Apple to tighten macOS's Full Disk Access controls for every AI agent, not just Meta's. Wonder and Wonder Full inherit that same agent product and that same trust gap -- a new model card's safety section doesn't retroactively settle a dispute about what the old one did.

Meta's own announcement, notably, lists no consumer or developer pricing for Wonder Full at all -- a point TechCrunch, Reuters (via Global Banking & Finance) and Meta's own post all agree on. One outlet, SQ Magazine, reports a specific figure anyway: $50 or $200 a month through Muse Code, attributed to the article's own reporting rather than a quoted Meta source. No other outlet checked here corroborates that number.

Markets read the launch as a win regardless of the access footnotes: Meta shares rose roughly 4% in the hours after the announcement, per Reuters. That reaction is about the signal -- Meta has a model independently measured ahead of two frontier rivals, for the first time -- more than the fine print about who is actually allowed to call it. The fine print is still the story for any developer deciding which Muse model to build on: the one the API will sell you is the one Meta's own marketing described accurately; the one that would have made the stronger pitch is not for sale.