Meta ships Muse Wonder and Wonder Full, its first frontier AI models

Meta enters the frontier race

Meta's Superintelligence Lab has shipped its first frontier-class models, a pair of systems called Muse Wonder and Muse Wonder Full that the company presents as state of the art. According to TechCrunch, the release is Meta's first attempt at a frontier model, and its first since Scale AI founder Alexandr Wang reorganised the company's AI efforts in August 2025.

A step up from Muse Spark

Meta's earlier model, Muse Spark, was built to be small and fast, trading size for low latency. The Wonder models take the opposite approach: TechCrunch reports they are larger and more capable, with stronger benchmark results and better performance on demanding workloads such as coding.

In the blog post announcing the launch, Meta described the pair as "the next step on our scaling ladder and a major leap toward personal superintelligence."

Since Wang's restructuring, Meta had confined itself to the lightweight Muse Spark model and its Muse agent, steering clear of the flagship releases that OpenAI and Anthropic are known for. Wonder changes that posture.

Benchmarks, pricing and availability

In testing detailed in the model card, Meta says both Wonder variants score on par with Fable 5.1 and GPT-6 Astra. Those figures are the company's own rather than independent evaluations, so the parity claim currently rests on Meta's reporting.

The pricing matches the flagship positioning. Served through the Meta Model API, Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens, which TechCrunch notes places it roughly in line with GPT-6 Sol and Opus 5.5.

The two variants are also distributed differently. Muse Wonder Full will not be served through the API at all; it is exclusive to subscribers of Muse Code. Both models will additionally power Meta's Muse agent.

Memory as the differentiator

The clearest strategic signal in the launch is what Meta calls "active memory" — the models' ability to retain and synthesise information a user shares over time. The emphasis fits Meta's stated interest in personal assistants that can carry out tasks over extended periods.

To support the claim, Meta pointed to MemoryBench, a benchmark aimed at continual learning and memory construction frameworks. According to the company's testing, Muse Wonder now scores above Fable, Opus and Astra on that measure.

"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," research scientist Kavya Kopparapu said in a video accompanying the launch, as reported by TechCrunch.

TechCrunch also notes that the models were trained at Meta's Prometheus data center in central Ohio, a facility that came online earlier this year.

Why it matters

For more than a year, Meta's AI arm had no direct answer to the frontier models coming from OpenAI and Anthropic. With Muse Wonder, it is arguing — on the strength of self-reported benchmarks, at least — that it has rejoined the top tier.

The more interesting bet is the differentiator. Rather than claiming outright leadership on general benchmarks, Meta is leaning into persistent personal memory as the capability that will decide the assistant era. If retaining a user's context across weeks and months turns out to matter more than a point or two on coding evaluations, Meta's MemoryBench lead could prove decisive.

The API pricing also signals intent: by matching the cost of rival flagship models, Meta is positioning Wonder to compete for developer workloads as well as its own agent products, not merely to serve as an in-house engine.