Meta's Superintelligence Lab unveiled Muse Wonder and Wonder Full, state-of-the-art models matching GPT-6 and Fable 5.1 performance with focus on long-term memory.
By Adam Makins·
Meta’s Superintelligence Lab just dropped a pair of new models that signal the company is finally ready to compete in the frontier AI space. Muse Wonder and Muse Wonder Full represent Meta’s first real attempt at building large-scale, highly capable AI models that can go toe-to-toe with OpenAI’s latest offerings.
These aren’t incremental updates. According to Meta’s own testing, Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra. That’s a massive jump from the company’s previous focus on lightweight models like Muse Spark, which prioritized speed and efficiency over raw capability.
The shift in strategy is intentional. Where Spark was designed for low-latency performance, the Wonder models are larger and built to handle complex tasks like coding and reasoning that require more computational muscle. Meta is being explicit about its ambition here, describing these models as marking “the next step on our scaling ladder and a major leap toward personal superintelligence.”
What’s particularly interesting about Wonder is Meta’s emphasis on something they’re calling “active memory.” This isn’t just about recalling information; it’s about a model that can retain and synthesize user-provided information over extended periods. Think of it as the difference between a chatbot that forgets your preferences after each conversation and an AI assistant that genuinely learns about you over weeks and months.
To test this capability, Meta highlighted a benchmark called MemoryBench, which challenges continual learning and memory construction. According to the company’s results, Muse Wonder now outperforms Fable, Opus, and Astra on this metric. For anyone building personal AI assistants, this is a meaningful differentiator.
Costs matter in the tech world. Through the Meta Model API, Wonder will run you $4 per million input tokens and $20 per million output tokens. That puts it roughly on par with GPT-6 Sol or Opus 5.5 pricing-wise, which is fair given the performance level.
There’s one caveat: Wonder Full, the more powerful variant, won’t be available through the API. It’s exclusive to Muse Code subscribers and will primarily power Meta’s Muse agent. This is a smart move that lets Meta control the experience while maintaining tier separation.
This release arrives at a critical moment for Meta’s AI strategy. Since Alexandr Wang restructured Meta’s AI efforts in August of last year, the company has been relatively quiet on frontier models. They released Muse Spark and the Muse agent, but nothing that directly challenged the state-of-the-art models OpenAI and Anthropic keep pushing out.
Wonder changes that calculus. It signals that Meta isn’t content being a player in the middleware game. They want competitive models at the highest level.
Worth noting: Meta trained these at Prometheus, its data center in central Ohio that came online earlier this year. That’s infrastructure investment paying dividends.
The frontier model space is getting crowded, and that’s good for users. More competition typically drives innovation, better benchmarks, and clearer thinking about what these models actually do well versus where they still fall short. Meta bringing serious contenders to the table adds pressure on everyone else to keep improving.
The big question now isn’t whether Meta can build capable models. They clearly can. It’s whether they can build the ecosystem and distribution channels to make these models matter in real applications.
Source: Infeeds
