Meta Enters the Frontier Model Race with the Launch of Muse Wonder

For a long time, Meta stayed out of direct competition in the frontier model category led by companies such as OpenAI and Anthropic. That has now changed: the company's Superintelligence Lab has unveiled two new models, Muse Wonder and Muse Wonder Full, which Meta presents as its first serious entry into this category. With them, the company is betting on advanced capabilities in coding and complex tasks, as well as a feature it calls "active memory."

From Lightweight and Fast to Powerful and Capable

The previous model, Muse Spark, was lightweight and designed for fast, low-latency responses. The Wonder family takes a different direction. These are larger, more capable models that, according to Meta, deliver better results on benchmarks and on tasks requiring multi-step reasoning, chief among them writing code.

Benchmark results published by Meta in the model card indicate that Wonder and Wonder Full perform comparably to Fable 5.1 and GPT-6 Astra. Note that these figures come from Meta itself, so independent verification is still needed to assess where the two models actually stand among their competitors.

The company describes the launch as the next step in scaling its models and a major leap toward what it calls "personal superintelligence."

Pricing and Access

The models' larger size is also reflected in their pricing. Meta has set the cost of using Muse Wonder through the Meta Model API as follows:

• $4 per million input tokens.
• $20 per million output tokens.

This pricing puts the model close to GPT-6 Sol and Opus 5.5, in the tier of advanced models aimed at professional use.

Wonder Full, however, will not be available through the API and will be exclusive to Muse Code subscribers. Meta appears to want its most powerful version to serve as a competitive advantage for its own products rather than making it available to all developers. Both models will also be used together to power Muse Agent, Meta's task-executing assistant.

Active Memory: Meta's Biggest Bet

Meta has long sought to build personal assistants that work with users over extended periods, not just in isolated conversations. For this reason, it focused the new models on their ability to retain information provided by the user and connect it, a capability the company calls "active memory."

To measure this capability, the company relied on MemoryBench, a benchmark designed to test continual learning and memory-building frameworks. Meta says Muse Wonder outperformed the Fable, Opus, and Astra models on this benchmark.

In a video accompanying the launch, researcher Kavya Kopparapu explained that the team put significant effort into enabling the model to remember what matters to the user over days, weeks, months, and beyond.

If this capability holds up in real-world use, it could offer clear practical value, including:

• Following long-running projects without having to re-explain the context each time.
• Providing more personalized assistance based on the user's accumulated preferences.
• Supporting AI agents in carrying out tasks that span long periods of time.

The First Fruits of the Restructuring

This is the first frontier model Meta has released since Alexandr Wang, founder of Scale AI, restructured the company's AI efforts in August of last year. Since then, Meta had released only the lightweight Muse Spark model and Muse Agent, and had held back from the kind of advanced models OpenAI and Anthropic are known for.

The two models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year.

What Does This Mean?

The Muse Wonder family marks a shift in Meta's strategy, from a focus on lightweight, fast models to direct competition at the top of the language model market. The final verdict will depend on hands-on experience and independent evaluations, but the company's focus on long-term memory gives it a distinctive angle that could prove influential in the race to build intelligent personal assistants.

✦ بقلم فريق دروب أيديا