Meta's Superintelligence Lab has unveiled two new large-scale AI models - Muse Wonder and Muse Wonder Full - marking the company's first serious entry into the frontier model category. The announcement came on Monday and signals a significant shift in ambition for Meta's AI division, which has until now focused primarily on lighter, faster models rather than competing directly at the top tier of AI capability.

Unlike the previously released Muse Spark, which was designed to operate with minimal latency and a smaller computational footprint, the Wonder models are built for raw capability. They are larger in scale, achieve stronger results on industry benchmarks, and are said to handle complex tasks - including code generation - with notably greater proficiency.

Benchmark Performance and Competitive Positioning

According to internal evaluations published in the model card, both Muse Wonder and Muse Wonder Full deliver benchmark scores that are broadly comparable to those of Fable 5.1 and GPT-6 Astra. Meta framed the release as a pivotal moment in its development roadmap, describing the models as "the next step on our scaling ladder and a major leap toward personal superintelligence" in an official blog post accompanying the launch.

The pricing structure for API access reflects the models' position in the high-performance tier. Developers accessing Muse Wonder through the Meta Model API will be charged $4 per million input tokens and $20 per million output tokens - a cost profile that puts it roughly on par with GPT-6 Sol and Opus 5.5, two of the more capable commercially available models from competing labs.

Muse Wonder Full, however, will not be available through the API at all. Access to the more powerful variant will be restricted exclusively to Muse Code subscribers. Both models will also serve as the underlying engine for Meta's Muse agent product.

A Focus on Long-Term Memory

One of the most prominently featured capabilities in the launch is what Meta refers to as "active memory" - the ability of the models to retain, organize, and draw upon information that users have shared over extended periods. This emphasis aligns with Meta's broader strategic interest in developing AI-powered personal assistants that can maintain continuity across conversations and tasks spanning days, weeks, or even months.

To substantiate these memory claims, Meta pointed to performance on MemoryBench, a benchmark specifically designed to evaluate continual learning and the construction of long-term memory frameworks in AI systems. According to Meta's own testing, Muse Wonder now outperforms Fable, Opus, and Astra on this benchmark - a result the company presented as a key differentiator for the new models.

"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 Kavya Kopparapu, a research scientist at Meta, in a video released alongside the launch.

The active memory capability is particularly relevant to Meta's vision of AI that functions as a persistent personal assistant rather than a stateless tool that forgets each conversation once it ends. By building long-term recall directly into the model architecture, Meta is positioning Muse Wonder as something closer to a continuous collaborator than a one-off query engine.

Context: A Restructured AI Division

The release of Muse Wonder and Muse Wonder Full represents the first frontier-level model launch from Meta since Scale AI founder Alexandr Wang took over the restructuring of Meta's AI operations in August of last year. That reorganization led to the creation of Meta's Superintelligence Lab and set a new strategic direction for the company's AI efforts.

In the period following the restructure, Meta released the lightweight Muse Spark model and introduced the Muse agent, but the company had notably stayed away from the kind of high-profile, state-of-the-art model launches that competitors like OpenAI and Anthropic have become known for. Monday's announcement changes that posture, with Meta now directly competing in the frontier model space for the first time under its restructured leadership.

The new models were trained at Meta's Prometheus data center located in central Ohio, a facility that came online earlier this year. The use of a dedicated, large-scale data center for training reflects the substantial infrastructure investment required to develop models at this level of capability, and underscores Meta's commitment to building out its own compute capacity rather than relying entirely on third-party cloud providers.

What Comes Next

With the Wonder models now publicly available - at least in part - Meta has entered a competitive segment of the AI market that it had previously sidestepped. The company's decision to restrict Muse Wonder Full to Muse Code subscribers while keeping Muse Wonder accessible via API suggests a tiered strategy aimed at both developer adoption and deeper integration with its own product ecosystem.

The emphasis on memory, long-horizon task performance, and personal assistant functionality also points toward where Meta sees the next frontier of AI utility - not just in answering questions, but in maintaining an ongoing, evolving understanding of the people who use these systems. Whether that vision translates into meaningful real-world adoption will likely depend on how well the active memory features perform outside of controlled benchmark conditions.

For now, the launch of Muse Wonder marks a clear inflection point for Meta's AI ambitions - one that moves the company from the periphery of the frontier model conversation to a direct participant in it.