Muse Wonder Full sits just below Claude Opus 5.5 on the Artificial Analysis index, Meta says

A platform lead handed a ticket on Monday to add Meta to the model bake-off starts with a thin file: a blog post, two news write-ups and benchmark results that Meta reports. The Verge notes that Meta pointed to results from internal testing on several industry benchmarks ^[5]. Wired concluded that the release suggests Meta is competitive again ^[12], and in the same piece wrote that benchmarks are not always a reliable measure of real-world performance ^[7].

The results Meta cites fall into two groups. The index score is a general ranking ^[1]. The legal-document and software results are narrow ^[6], and in our view a team that does neither kind of work finds little in them. Anyone whose shortlist includes Google has a further gap, since The Verge wrote that it is unclear how the models stack up against Google's Gemini 4 Argon ^[16].

Meta's own summary comes from its blog post. The company said, "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence." ^[13] It is roadmap language. The sentence holds a ladder and a destination and no unit a buyer can test.

Active Memory and Active Focus are the part a team can test directly. Meta designed them to keep the model locked onto a stated goal ^[10]. The failure they target is expensive on long jobs, where confusion in complex research or a large software build can lead to compounded errors ^[11]. A test for that looks different from a leaderboard run. It takes a task of many steps in the team's own repository or document set and records where the model drifts from the goal. Meta did not disclose whether it has seen agents powered by Muse Wonder misbehave in testing ^[17].

Continuity is a separate matter. Wired reports that Zuckerberg put billions into Meta Superintelligence Labs and hired Alexandr Wang to lead it, and that the effort has sometimes been rocky, with some researchers not staying long ^[15]. A team weighing a long contract can reasonably keep the first one short.

Teams often tell themselves that a high index position predicts how a model will do on their own work. The sources support a narrower use of the number: deciding whether the model gets tested at all. Two questions sort that decision. The first is whether the team's work resembles the legal-document and software tasks Meta cites. The second is whether the team has confirmed in writing which model its production route serves, since the two reports describe access differently ^[2][4].

Where both answers are yes, Full belongs in a bake-off against the incumbent, scored on drift from the stated goal as well as on the final answer. A matching workload on an unconfirmed route waits for Meta's reply, and a confirmed route with a different workload gets a test only if the test is cheap. Two no answers leave Meta off this quarter's list.