Published fact: Meta Research published Muse Wonder: A Frontier Model For Personal & Professional AI Agents under a byline reading “October 12, 2026”. This article is written two days later, on October 14, 2026, and works from that page alone.
The page announces a series, not a model, and gives it three destinations at once. That is where the useful question sits: if most of the argument the page makes is carried by Muse Wonder Full, why does the self-serve API route get plain Muse Wonder instead, and what should a team reading this page conclude about which of the two it can actually plan around?
Thesis, and it is mine: the product decision is not in any capability claim. It is in the availability sentence. Meta has split one announcement across two access tiers that point at two different audiences, and the model the evaluation section leans on is the one a developer cannot reach through the published API channel. Everything else on the page is framing around that split.
What the page releases
Published fact: the two names are introduced together and described in the same breath — “Today, we’re excited to introduce the Muse Wonder series with Muse Wonder and Muse Wonder Full, our most capable models that deliver state-of-the-art performance on personal and professional agentic tasks.”
Published fact: the availability statement names three channels in one sentence — “Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full is available today for Muse Code subscribers.” The page links “Meta Model API” and “Muse Code” to developer destinations; it does not link, price or define what a Muse Code subscriber subscribes to.
Published fact: the two capabilities the page builds the release around are Active Memory and Active Focus. Active Memory is described as “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand”. Active Focus is described as how the model keeps working “toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information”.
Published fact: the page’s summary of its own results is that “The Muse Wonder series achieves state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding.” It directs readers elsewhere for the substance: “For more details about our evaluations, see our report.”
Published fact: the training-site statement is “Trained at our gigawatt-scale Prometheus data center”, and it credits “teams across Meta”.
Vendor claim: the ladder metaphor, the framing of this as “delivering our most capable models yet”, the description of the pair as “the next step on our scaling ladder and a major leap toward personal superintelligence”, and the fourteen-month claim that “Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts.” None of these is checkable from the page.
Where the two names separate
Published fact: the page never separates them on capability, size or price. There is no per-model specification block, no second availability sentence, and no statement of how far apart the two models are in quality.
Inference, mine: for a reader, that means the difference between the two names, as published, is a distribution channel and nothing else. A capability gap may well exist — the page implies one — but the page leaves it as an implication.
Which model the evaluation carries
Published fact: the scatter analysis is run on Muse Wonder Full, not on plain Muse Wonder. The page states it directly: “The plot below compares Muse Spark 1.3 and Muse Wonder Full, with each point denoting performance at a distinct reasoning effort.”
Vendor claim, with the page’s own stated conditions: “Both of our models improve similarly on out-of-distribution tasks as we scale test-time compute. But given its in-distribution performance, Muse Wonder Full performs even better on unfamiliar tasks than the Muse Spark 1.3 trend would predict.” The page’s mechanism for that follows in the same paragraph: “This upward shift is because Muse Wonder and Muse Wonder Full build on larger, stronger base models: a stronger base yields post-training gains that carry over more effectively to unfamiliar tasks.”
Published fact: the page’s own statement of what its results cover is that “The full results below compare Muse Wonder and Muse Wonder Full with other frontier models across a broad evaluation suite”, and its own scoping sentence explains the four areas it then treats in depth.
Inference, mine, about the page’s text layer: it contains no numeric score at all. Benchmark values live in chart images, whose captions name the evaluations but carry no figures, so “state-of-the-art” cannot be read as any particular margin, and the gap between Muse Wonder and Muse Wonder Full cannot be measured from the page.
The third channel is a plan
Published fact: the consumer destination is written in the future tense throughout. “Muse Wonder will soon power experiences in Muse, and will roll out first to subscribers.” Then: “Muse Wonder will also power artifact generation and processes that require more intelligence.” Then: “And in the future, we’ll be using Muse Wonder to develop specialized models that we can bring to everyone using Muse.”
Inference, mine: three forward-looking statements, no shipped consumer product. Note which model is promised there — plain Muse Wonder again, not Muse Wonder Full.
What the page withholds
The gaps are consistent: no price for either model, for Muse Code or for the API; no context window; no parameter count or architecture description; no statement of which regions or markets any channel reaches; no model identifier, endpoint or migration note for Muse Spark users; no open-weight, licence or redistribution statement; and no definition of a Muse Code subscriber or what that tier includes. Missing above all is the one number a buyer would need — how much better Muse Wonder Full is than the Muse Wonder the API serves.
How to read the ordering
Published fact supplies two names, one availability sentence, a stated capability set and vendor-reported results with no figures attached. Vendor claim supplies the ladder framing. Inference, mine: Meta has published the version it wants developers to adopt into the open channel and kept the version its own evaluation section argues about behind a subscription, while the consumer surface stays unbuilt. A team reasoning from this page is choosing between the API sibling now and an upgrade path whose size is unpublished.
What would unsettle that reading: per-model numbers showing the two are close, a definition of Muse Code subscribers broad enough to cover most teams intending to build, or a capability statement that ties each name to a distinct job. None of those is on the page.