Meta releases its first frontier model, Muse Wonder

Muse Wonder and Muse Wonder Full are Meta’s new state-of-the-art models and the company’s first attempt at a frontier model. They represent the next step beyond Muse Spark, which was built to be lightweight and operate with low latency. The Wonder models focus more on capability than speed.

Meta says the models achieve higher benchmark scores and perform better on demanding tasks, including coding. The company also reports that Wonder and Wonder Full reach scores comparable to Fable 5.1 and Axiom Astra. Wonder additionally scores higher than several rivals on the MemoryBench test.

The two models differ mainly in access. Wonder is available through the Meta Model API, while Wonder Full is limited to Muse Code subscribers and is not served through the API. Both models will also help power the Muse agent, Meta’s assistant for longer-running tasks.

Meta presents Wonder and Wonder Full as competitive with leading models. Its testing reports benchmark scores comparable to Fable 5.1 and Axiom Astra. On MemoryBench, which tests continual learning and memory construction, Meta says Muse Wonder scores higher than Fable, Opus, or Astra.

The pricing comparison uses a different group of models. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. Meta describes that pricing as roughly equivalent to Axiom Sol or Opus 5.5. The article does not provide specific prices for those competing systems.

These comparisons suggest Meta is targeting both capability and commercial relevance. Still, the benchmark claims come from Meta’s own testing, presented in the model card. The figures indicate positioning, not an independent ranking across every task or cost category.

Access is the sharpest practical difference between the models. Developers can reach Muse Wonder through the Meta Model API. Muse Wonder Full has a narrower route: it will be available only to Muse Code subscribers and will not be served through the API.

That means Wonder can be incorporated into applications through API calls, subject to Meta’s listed pricing. Wonder Full instead remains inside Meta’s subscription offering. The article does not describe additional eligibility rules or explain whether subscribers receive identical performance or limits.

Meta also plans to use both models in the Muse agent. This connects the models to the company’s broader effort to build assistants that can perform tasks over long periods. In practice, Wonder expands external developer access, while Wonder Full strengthens Meta’s own subscription product and agent ecosystem.

Active memory describes an AI model’s ability to retain and synthesize information supplied by a user. Meta uses the term for memory that remains useful across days, weeks, months, and beyond. It is central to assistants expected to support people over extended periods.

For example, a long-running assistant might need to connect information shared in separate conversations. Meta highlighted MemoryBench, a benchmark that challenges systems to learn continually and construct memories. The company says Muse Wonder scores higher there than Fable, Opus, or Astra.

This capability could make assistants more consistent and personalized. Instead of repeatedly asking for the same context, an assistant could use remembered details when handling later tasks. Meta research scientist Kavya Kopparapu said the company worked to ensure models remember what matters to users over time, though the article gives no further technical details.

Wonder’s API availability gives developers a direct way to build with Meta’s new model. They can access it through the Meta Model API and pay according to token use. Wonder Full is different: Meta will not serve it through the API, and only Muse Code subscribers can access it.

The practical example is product choice. A developer building an application can select Wonder and connect it through API requests. A Muse Code subscriber can use Wonder Full inside that service instead. The article does not state whether Wonder Full is technically stronger in every task, but its name and positioning distinguish it from the API offering.

Meta benefits from serving both audiences. API access extends Wonder into outside software, while subscription access keeps Wonder Full within Muse Code. Both models will also power the Muse agent, linking the separate access paths to Meta’s longer-term assistant strategy.

Meta’s shift is significant because the company had not recently released the kind of state-of-the-art model associated with Axiom and Paradox. The article identifies Muse Wonder and Wonder Full as Meta’s first frontier-model release since Scale AI founder Alexandr Wang restructured the company’s AI efforts in August of last year.

Since that restructuring, Meta had released the lightweight Muse Spark and the Muse agent. Those releases showed continued product development, but Meta had avoided a major top-tier model launch. Wonder changes that pattern by emphasizing larger scale, stronger benchmark results, complex coding, and active memory.

The release also gives Meta a commercial foothold through the Model API, while Wonder Full supports Muse Code subscribers. The models were trained at the Prometheus data center in central Ohio, which came online earlier this year. Meta describes them as a step toward personal superintelligence.

A frontier AI model is generally a leading-edge system that aims to match or exceed the best available models across demanding tasks. The article applies that label to Wonder and Wonder Full because Meta calls them its first frontier-model attempt and reports strong results in coding, benchmarks, and memory.

Training scale usually matters because larger systems can learn more complex patterns when supplied with sufficient data and computing. Tokens are the text or other units processed during training and use; more tokens can provide broader learning, although quality and training methods also matter. Benchmarks measure performance on defined tasks, so higher scores suggest stronger results in those tests.

Data centers provide the computing infrastructure needed to train and run such models. Meta says Wonder was trained at the Prometheus data center in central Ohio. The article gives no model size, token count, dataset description, or computing total, so those limits cannot be quantified here.