Meta releases Muse Wonder, its first frontier AI model

Meta released its first frontier AI models on Monday, Muse Wonder and Muse Wonder Full, and claims they achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra. The release marks the company's entry into the top tier of AI model development after more than a year in which rivals OpenAI and Anthropic dominated state-of-the-art releases.

The models come from Meta's Superintelligence Lab. According to Meta's testing, presented in the model card, Wonder and Wonder Full post higher benchmark scores and better results on complex tasks such as coding than the company's earlier lightweight model, Muse Spark. Meta positioned Muse Spark as a low-latency model; the Wonder models are larger and more capable.

The stakes are straightforward. Frontier models anchor the API business, the agent products and the enterprise contracts that define the current AI market, and Meta has not competed at that level until now. The Wonder release signals that the company intends to.

What does the release actually include?

Two models, with sharply different distribution:

• Muse Wonder — available through the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens. Meta's blog post describes this pricing as roughly equivalent in cost to GPT-6 Sol or Opus 5.5.
• Muse Wonder Full — available only to Muse Code subscribers and not served through the API at all.
• Both models will also power the Muse agent, Meta's existing assistant product.

The pricing places Wonder in direct commercial competition with the premium tiers of OpenAI's and Anthropic's lineups, the same segment its benchmarks target.

What does Meta claim for performance?

Meta's own testing, laid out in the model card, puts Wonder and Wonder Full at benchmark parity with Fable 5.1 and GPT-6 Astra. The company reports stronger results on complex tasks, including coding, than its lightweight predecessor.

The blog post frames the release in expansive terms. "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," it reads.

Independent verification of the benchmark claims was not part of Monday's announcement; the figures come from Meta's internal evaluations. That distinction matters for buyers comparing API costs against measured performance, since model-card benchmarks routinely reflect the testing choices of the lab that produced them.

What is "active memory" and why did Meta emphasize it?

The most distinctive claim in the release concerns memory. Meta placed particular emphasis on the models' ability to retain and synthesize information supplied by the user over long periods, a property the company calls "active memory." The emphasis fits Meta's stated interest in personal assistants that can perform tasks over a long duration rather than answer single queries.

To measure recall, Meta highlighted MemoryBench, a benchmark that tests continual learning and memory construction frameworks. According to the company's testing, Muse Wonder now scores higher on MemoryBench than Fable, Opus or Astra.

Research scientist Kavya Kopparapu, in a video accompanying the launch, said: "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."

The claim carries commercial weight. Persistent, user-specific memory is the technical underpinning of any assistant that manages schedules, projects or preferences across months, and Meta has tied its superintelligence ambitions to exactly that product form. If active memory holds up under third-party testing, it differentiates Wonder from competitors scored on the same benchmark.

Why is this Meta's first frontier model?

The Wonder models are the first frontier release since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year. Under that restructured organization, Meta shipped the lightweight Muse Spark model and the Muse agent but avoided the state-of-the-art releases that OpenAI and Anthropic are known for.

Monday's announcement ends that gap. Meta trained the models at its Prometheus data center in central Ohio, which came online earlier this year, according to the post. The facility gives Meta dedicated compute for frontier-scale training, and the Wonder release is the first visible output of that infrastructure.

The sequence — restructure, data center, lightweight model, agent, then frontier model — reads as a deliberate scaling path rather than a rushed catch-up. Meta's blog post calls the Wonder models "the next step on our scaling ladder," which implies further steps above them.

What happens next?

Two questions will determine whether the release matters beyond the announcement. First, whether third-party evaluations confirm the claimed parity with Fable 5.1 and GPT-6 Astra, particularly on coding and the MemoryBench recall tasks Meta highlighted. Second, whether developers accept the $4.00/$20.00 per-million-token pricing against the models it is priced alongside, given that the larger Wonder Full variant is locked to Muse Code subscribers and absent from the API entirely.

The answer will shape the next round of Meta's scaling ladder, which the company has already said does not stop here.