Meta Claims Muse Wonder Beats Claude Fable 5.1 — Without Independent Verification
With Muse Wonder and Muse Wonder Full, Meta claims to be back among the world's best language models, just weeks after the company delayed the launch over safety concerns. But the company does not disclose how the models were built, and no independent evaluation is available.
Meta launched its newest AI models, Muse Wonder and Muse Wonder Full, on October 12, 2026. The company presents them as a breakthrough: according to Meta's own blog post, as relayed by WIRED, the largest models are meant to measure up to the best from Anthropic and OpenAI. "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," the company writes in the blog post, cited by WIRED.
The launch arrives in a deliberately charged context. Last month, Mark Zuckerberg said in a post on X that the company had delayed the launch of the Muse models in order to work on safety. Earlier in October, Meta also published a framework intended to ensure the company scales the models' capabilities safely. Muse Wonder is thus the first concrete deliverable since Meta itself made safety a precondition for progress — which makes both the capability and safety claims particularly worth scrutinizing.
What Meta claims about performance
Meta's benchmark figures come from third-party indexes, but it is important to emphasize that the comparisons themselves are the company's own assertions, relayed through WIRED's coverage. According to Meta, the largest version of Muse Wonder surpasses Anthropic's second-best model, Claude Fable 5.1, and sits just below the company's best, Claude Opus 5.5, on the Artificial Analysis Intelligence Index — a widely used, aggregated score of models' capabilities.
In addition, Meta claims the model beats GPT-6 Astra from OpenAI and both Claude models on certain specialized benchmark tests. WIRED points out a well-known caveat at the same time: benchmark scores are not always a reliable indicator of how a model actually performs on real-world tasks. No independent evaluation of Muse Wonder is available at this stage, meaning readers must for now rely on Meta's own numbers.
Availability and use
The models are being made available via an API and through Muse Code, Meta's tool for AI-assisted software development. The top version will also power features across Meta's platforms, including the personal agent Muse. That means the model is not only a developer offering, but also the underpinning of Meta's consumer-facing agent strategy.
Active Memory and Active Focus: new controls for agents
The most concrete novelty in the launch is two features aimed at agent reliability: Active Memory and Active Focus. According to Meta, they are designed to keep the model locked to a stated goal and prevent it from becoming confused — a well-known problem for AI agents, where small errors early in a multi-step task can compound and cause the agent to go completely off track.
WIRED's headline points precisely at this: the model is to include "new features that keep agents from going off track." What the features involve technically, and how effective they are, is not specified in the available material. For now, it is Meta's own description of their purpose that exists — not documentation of results.
The safety claims — and the context around them
The safety portion of the launch is closely tied to Zuckerberg's delayed-launch post and the new safe-scaling framework from earlier in October. Meta says the company has put significant work into training Muse Wonder to refuse to generate dangerous answers, such as recipes for biological weapons or assistance with hacking, and that the model scores well on several benchmark tests designed to measure how resilient models are against such attempts.
The same caveat applies here as well: these are the company's own statements, and there is no independent verification. Given that Meta itself used safety as the justification for delaying the launch, it is a natural follow-up question how much concrete documentation the company will present for these claims.
What is missing
The biggest open questions after the launch concern information Meta has not provided. According to WIRED, the company did not disclose details about how the new Muse Wonder models were built or trained. It is also not known — from Meta or anyone else — whether agents powered by the models misbehaved during testing, which is especially relevant given that agent reliability is a central selling point of the launch.
Taken together, Muse Wonder emerges as a clear signal that Meta wants back at the top of the AI competition, and that the company is now selling safety and agent control as part of the product. But with its own benchmark numbers, its own safety claims, and no independent evaluation, it is for now the company's word that remains — not documented competitiveness. The coming weeks will show whether independent testing and closer technical documentation confirm the picture Meta paints.