Meta releases Muse Wonder, its first attempt at a frontier AI model

Meta released its first genuine frontier-class artificial intelligence model on Monday, a pair of systems called Muse Wonder and Muse Wonder Full that the company is positioning as its answer to OpenAI's and Anthropic's top-tier models after more than a year of public struggle to keep pace in the AI race.

The release, detailed in a post on Meta's research blog, marks a shift from the company's earlier Muse Spark models, which Meta had built for speed and low cost rather than raw capability. Wonder and Wonder Full are larger systems that Meta says close the gap with rival frontier models on coding, professional-agent tasks and what the company calls "active memory," the ability of a model to retain and reuse information a user has shared in earlier conversations.

The numbers

Muse Wonder is available immediately through the Meta Model API at $4 per million input tokens and $20 per million output tokens, pricing Meta and outside analysts describe as roughly in line with competing frontier systems from OpenAI and Anthropic, according to TechCrunch's report on the launch. Muse Wonder Full, the larger of the two systems, is reserved for subscribers to Meta's Muse Code product and is not sold through the public API.

Both models were trained at Prometheus, Meta's roughly 1-gigawatt data-center supercluster in New Albany, Ohio, which came online this year. On memory retention, Meta cites its own MemoryBench evaluations, a benchmark described in a paper on continual learning in large language model systems, and says Wonder outperforms named rival models on that test. The company's published model card adds that the models carry Meta's strongest safeguards to date against misuse involving chemical, biological, radiological, nuclear and cyber risks, though it does not disclose the underlying numeric benchmark scores in the accompanying text, leaving most of that detail to charts.

How Meta got here

The launch caps roughly fourteen months of reorganization inside Meta's AI division. The company brought in Alexandr Wang, formerly chief executive of Scale AI, in the summer of 2025 to lead a new Meta Superintelligence Labs unit after Meta's open-weight Llama models lost ground to closed competitors. That reshuffle split the AI group into separate teams for frontier models, fundamental research, infrastructure and products, a reorganization that was itself revisited within months, as Meta tried several structures in quick succession.

Since then, Meta has released a string of smaller products, including Muse Spark in April, an open-weight Muse Glimmer model in August, and the Muse personal assistant in September, but had avoided a head-to-head release against the state-of-the-art systems fielded by OpenAI and Anthropic. Muse Wonder is the first model Meta has explicitly framed as a frontier contender rather than a lightweight or open-source offering.

Meta's decision not to release Wonder's weights openly also continues a break from the company's earlier Llama strategy, which built much of Meta's standing in the AI research community on freely downloadable models. Wang has said in previous public comments that Meta intends to open-source future versions of its flagship models, but no timeline has been set, and Wonder, like Muse Spark before it, remains available only through a paid API and subscription products. That approach puts Meta in the same closed-model camp as OpenAI and Anthropic for its most capable systems, even as it continues to release smaller open-weight models, such as Muse Glimmer, further down its product line.

The Prometheus data center where Wonder was trained has itself been a point of scrutiny in Ohio, where local officials have highlighted the facility's reliance on on-site natural gas generation and new nuclear power agreements to meet its electricity demands, even as some residents raised concerns about strain on the regional grid. Meta has described Prometheus as one of several multi-gigawatt "titan clusters" it is building to support the computing needs behind models like Wonder.

Who is affected

Developers building on the Meta Model API now have access to a more capable, more expensive model alongside the cheaper Muse Spark line. Subscribers to Muse Code, Meta's coding-focused subscription, get exclusive access to the larger Wonder Full model. Meta also said both models will soon power its Muse consumer agent, rolling out first to paying subscribers before a broader release, meaning ordinary users of Meta's apps are likely to encounter the new systems indirectly through features such as automated trip planning, customer-service style tasks and in-app content generation.

The release is also being watched closely by OpenAI and Anthropic, against whose named models, GPT-6 Astra and Fable 5.1, Meta benchmarked its new systems. Meta's blog post and model card claim Wonder is competitive with both on coding and reasoning tasks and ahead of them on memory retention, though TechCrunch's report noted that those comparisons come from Meta's own testing and have not been independently verified.

Reaction

Meta framed the release in ambitious terms. In its research blog post announcing the models, the company wrote that Wonder represents a major step in its broader roadmap toward more capable, more autonomous AI systems.

"The next step on our scaling ladder and a major leap toward personal superintelligence," Meta said in the blog post announcing Muse Wonder and Muse Wonder Full.

Outside reaction so far has been more measured. Coverage of the launch has focused on whether Meta's self-reported benchmark numbers hold up once developers and independent evaluators can test the models directly, a recurring question for frontier-model launches across the industry. Meta has not yet published a full safety report for Wonder, saying in its model card that one is forthcoming.

What happens next

The near-term test for Muse Wonder will be adoption: whether developers paying $4 and $20 per million tokens choose it over established options from OpenAI and Anthropic, and whether independent benchmark trackers corroborate Meta's claims on coding and memory tasks. Meta said Wonder will roll out inside its Muse assistant in the coming weeks, starting with paying subscribers, which will give the company a much larger, if less technical, audience to judge the model's real-world performance.

Longer term, the release is a signal of Meta's intent to keep competing directly at the frontier rather than focusing solely on smaller, cheaper, or open-weight models, a strategy shift that follows more than a year of leadership changes inside its AI organization. Meta has also said it plans to use Wonder as a base for building specialized models for Muse users, suggesting the current release is a foundation for further products rather than a one-off launch. Whether Wonder closes the gap with rivals, or simply narrows it, will depend on results that outside researchers, rather than Meta's own charts, will need to confirm in the weeks ahead. A fuller safety report, which Meta has promised but not yet published, is likely to shape some of that independent scrutiny once it appears.