Hajra Naz
Meta’s Superintelligence Lab released two new models on Monday. They are Muse Wonder and Muse Wonder Full. The company describes them as state-of-the-art. They also mark Meta’s first attempt at a frontier model.
How Wonder Differs From Muse Spark
Muse Spark was built to be light and fast. It aimed for low latency. The Wonder models take a different path. They’re larger and said to be more capable. Meta says they score higher on benchmarks. They also do better on complex tasks like coding.
Benchmark Claims
Meta shared its test results in a model card. According to that testing, Wonder and Wonder Full reach scores comparable to Fable 5.1 and GPT-6 Astra. These are results from Meta’s own testing. Independent checks haven’t confirmed them yet.
“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta’s blog post reads.
Read More: Can Meta’s Muse Overcome Its Growing Trust Issues?
Pricing and Access
The larger scale shows up in the price. Developers can use Wonder through the Meta Model API. It costs $4 per million input tokens. Output tokens cost $20 per million. That makes it roughly equal in cost to GPT-6 Sol or Opus 5.5.
Wonder Full works differently. It’s only available to Muse Code subscribers. It won’t be served through the API. Meta will use both models to power the Muse agent.
The Focus on “Active Memory”
Meta wants to build personal assistants that handle tasks over long periods. So the company put special emphasis on memory. It says the new models can retain and combine information the user gives them. Meta calls this “active memory.”
Memory Benchmark Results
To show better recall, Meta pointed to MemoryBench. This benchmark tests continual learning and memory construction frameworks. According to Meta’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 launch video.
Why Memory Matters for Assistants
An assistant that forgets is frustrating. You repeat yourself. You re-explain your preferences. A model with strong memory can pick up where you left off. It can recall a past request or a stated preference. That makes it more useful over time.
Meta’s Frontier Strategy
These are the first frontier model releases since a major shake-up. Alexandr Wang, the founder of Scale AI, restructured Meta’s AI efforts in August of last year. Since then, Meta has released the lightweight Muse Spark and the Muse agent. It has avoided the kind of state-of-the-art releases that OpenAI and Anthropic are known for. Wonder changes that.
Read More: Research Reveals How GPT-6 Astra Created Fake Identities to Inject Code
Where the Models Were Trained
According to the post, Meta trained the models at its Prometheus data center in central Ohio. The facility came online earlier this year.
What to Watch Next
Several questions remain. Independent testing will show whether the benchmark claims hold up. Developers will weigh Wonder’s price against rivals at similar cost. Users will see how well “active memory” works in daily use.
