Meta Unveils First Frontier Models, Muse Wonder and Muse Wonder Full, Marking a Leap Toward Personal Superintelligence

A Milestone Release Amid Restructuring

On Monday, Meta’s Superintelligence Lab introduced its inaugural frontier models, Muse Wonder and the more powerful Muse Wonder Full, signaling the company’s entry into the high‑end AI arena after a period of focusing on lighter solutions.

The launch follows last August’s organizational overhaul led by Scale AI founder Alexandr Wang, which shifted Meta’s AI strategy toward more ambitious, large‑scale systems while still maintaining the previously released Muse Spark and the Muse agent.

Technical Edge and Benchmark Performance

Muse Wonder, built on a larger architecture than the earlier Muse Spark, delivers higher benchmark scores and excels on complex tasks such as coding, positioning it alongside competitors like Fable 5.1 and GPT‑6 Astra.

In Meta’s internal testing, both Wonder and Wonder Full achieved results comparable to those of Fable 5.1 and GPT‑6 Astra, according to the model card released with the announcement.

“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta’s blog post states.

Pricing Structure and Access

Developers can access Muse Wonder through the Meta Model API at a rate of $4 per million input tokens and $20 per million output tokens, placing its cost on par with models such as GPT‑6 Sol and Opus 5.5.

Muse Wonder Full, however, will be reserved for Muse Code subscribers and will not be available via the public API, ensuring a more controlled rollout for the most advanced capabilities.

Both models are also integrated into the Muse agent, extending their utility to personal assistance scenarios.

Active Memory and Long‑Term Retention

Meta emphasized the new models’ ability to retain and synthesize user information over extended periods, a feature the company terms “active memory.”

Performance on the MemoryBench benchmark, which evaluates continual learning and memory construction, shows Muse Wonder outperforming Fable, Opus, and 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.

Training Infrastructure

The models were trained at Meta’s Prometheus data center in central Ohio, which became operational earlier this year, providing the computational resources needed for such large‑scale AI development.

Overall, the release positions Meta more squarely in the frontier of AI research, offering developers a powerful, memory‑aware alternative to existing offerings from OpenAI and Anthropic while maintaining a clear pricing structure and strategic access controls.