Meta ships its first frontier model with Muse Wonder, betting on agents and active memory
Meta's Superintelligence Lab has moved from Muse Spark to a heavier, better-benchmarked model that powers its Muse agent and leans on a distinctive memory feature.
Meta has released its first model from the Superintelligence Lab: a pair called Muse Wonder and Muse Wonder Full. In a blog post posted on Monday, Meta described the pair as its first frontier model and the next step on its scaling ladder toward personal superintelligence.
This marks a real repositioning of the company's AI story. The two are the successors to Muse Spark, Meta's previous offering, and to earlier Llama releases, and they now power the company's Muse agent. The models are led, in Meta's own framing, by Superintelligence Lab chief Alexandr Wang, and Meta's marketing leans on the agent experience rather than a single model card.
What the pair actually is
The lab calls the Wonder models a frontier-grade agentic AI model pair, optimized for autonomy, tool calling, and active memory. Meta points to benchmark scores that it says are comparable to Fable 5.1 and GPT-6 Astra, and highlights the MemoryBench results, where Muse Wonder outperforms Fable, Opus, and Astra.
There are two tiers. Muse Wonder is the base model, available through the Meta Model API at $4 per million input tokens and $20 per million output tokens — a price point that sits at roughly the same cost-per-token as GPT-6 Sol and Opus 5.5, Meta says. Muse Wonder Full is the larger, more capable variant, and Meta restricts it to Muse Code subscribers; it is not exposed on the open API.
- First frontier model from Meta's Superintelligence Lab, the second release in the Muse line after Spark.
- Two tiers: Muse Wonder (open API) and Muse Wonder Full (Muse Code subscribers only).
- Agentic by design: optimized for autonomy, tool calling, and active memory that Meta says beats Fable, Opus, and Astra on MemoryBench.
- Priced to undercut rivals: $4 per million input tokens and $20 per million output tokens on the open model.
Why the memory angle leads
Meta leans hardest on the active memory feature. On its own MemoryBench, the company says Muse Wonder scores higher than GPT-6 Astra and OpenAI's Fable, and the framing is that Wonder keeps long threads of context rather than re-reading everything from scratch. For a model being sold as a personal agent, that is the differentiator: an assistant that remembers what you already told it, across sessions.
For developers, Meta now has a frontier model you can call at a competitive price, with a tier — Muse Wonder Full — aimed squarely at people who build on its coding products. For the rest of the market, it is another sign that competition in the frontier tier now runs on agents and memory rather than raw chat quality alone, and that one of the largest ad businesses in tech is seriously betting its stack on the path to personal superintelligence.
Sources: TechCrunch — Meta debuts Muse models from Superintelligence Lab, led by Alexandr Wang; The Verge — Meta's first superintelligence model: Muse Wonder; Meta Newsroom — Muse: Meta Superintelligence Labs' first release.
Corrections: Report an error at corrections@wireshaman.com.