Meta releases its first frontier model, Muse Wonder

Meta's Superintelligence Lab released a pair of new models on Monday, Muse Wonder and Muse Wonder Full, representing the company's first attempt at a frontier model. Where Muse Spark aimed to be a lightweight model that could perform at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks like coding.

According to Meta's testing, presented in the model card, Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra. Meta says the largest version of the new model outperforms Anthropic's second-best model, Claude Fable 5.1, and ranks just below its best, Claude Opus 5.5, on the Artificial Analysis Intelligence Index. Meta also says the model outperforms OpenAI's best model, GPT-6 Astra, and Claude Fable 5.1 on some specialised benchmarks, including ones that measure a model's ability to work with legal documents and to build software. Benchmarks are not always a reliable measure of real-world performance, but this is still a significant step forward for Meta's models.

"Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," Meta's blog post about the models reads.

That scale is reflected in the pricing. Accessed through the Meta Model API, Wonder will cost $4 per million input tokens and $20 per million output tokens, making it comparable in cost to leading frontier models. Notably, Wonder Full will only be available to Muse Code subscribers, whilst Wonder will be available through the Meta Model API and through Muse Code, a tool for writing software with AI. Both models will be used to power the Muse agent, with Meta saying the best version of the model will power features across its platforms, including its Muse personal agent.

In keeping with Meta's interest in developing personal assistants that can perform tasks over a long period, the company placed particular emphasis on the new models' ability to retain and synthesise information given by the user, a property Meta calls "active memory." To achieve better recall, Meta highlighted the MemoryBench benchmark, which challenges continual learning and memory construction frameworks. According to the company's testing, Muse Wonder now scores higher on MemoryBench than Fable, Opus, or Astra.

"We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," said research scientist Kavya Kopparapu in a video accompanying the launch.

Muse Wonder also includes two novel features designed to help make agents more effective and less inclined to misbehave. The new features, called Active Memory and Active Focus, are designed to keep the model locked into a stated goal and prevent it from becoming confused. This can be a challenge when AI performs complex research or builds a large software project, for example, and can lead to expensive mistakes or compounded errors.

Meta says it has put significant effort into training Muse Wonder to refuse to generate dangerous outputs, such as recipes for building bioweapons or help with hacking, and says the company scores well on several benchmarks designed to measure how resilient models are to those prompts. Zuckerberg said in a post on X last month that the company had delayed the release of its Muse models to work on safety. Meta also revealed a framework designed to ensure that it scales its models' capabilities safely in a blog post earlier this month.

The new models are the first frontier model release since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year. Since then, Meta has released its lightweight Muse Spark model and Muse agent but has avoided the kind of state-of-the-art model releases OpenAI and Anthropic are known for. The company has also released a small open-weight model called Muse Glimmer, and says it plans to release larger open-weight models in the future. It did not disclose details of how the new Muse Wonder models were built or trained.

According to the post, the models were trained at Meta's Prometheus data centre in central Ohio, which came online earlier this year. Meta was once a trailblazer in AI: its Llama series of models were the first frontier open-weight offerings from a big lab, and Llama, which could be downloaded and run on one's own hardware, was widely adopted by researchers and startups. The company fell behind after delivering several disappointing versions of the model, and as Anthropic, OpenAI, and Google rapidly introduced more capable models.

Muse Wonder may add more fuel to an already furious race to develop more capable AI.