Skip to content
NewsModel release

Meta ships its first frontier model, Muse Wonder, betting on memory over raw scale

· by Pondero Newsdesk

The short version

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, its first frontier-class models, priced at $4 per million input tokens with the stronger variant held back from the API.

Meta ships its first frontier model, Muse Wonder, betting on memory over raw scale

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, its first models the company calls "frontier," built around a feature it calls active memory rather than a straight benchmark-chasing scale-up, per Meta's model card.

What

Meta frames Wonder around two paired capabilities: Active Memory, which carries context about a user across conversations and retrieves the relevant details when a task needs them, and Active Focus, which the company describes as the model's ability to keep working on a hard goal, try alternate approaches, and check in with the user before taking important actions, per Meta's own announcement. "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," Meta research scientist Kavya Kopparapu said in a launch video, quoted by TechCrunch. On general benchmarks, Meta's own testing puts Wonder and Wonder Full at scores comparable to Fable 5.1 and GPT-6 Astra, according to the model card. On MemoryBench, a benchmark for continual-learning and memory-construction systems, and on FocusBench and SWE-Marathon, Meta's two measures of sustained task execution, the company reports Wonder and Wonder Full scoring ahead of Fable, Opus, and Astra, per the model card. Wonder is priced at $4 per million input tokens and $20 per million output tokens through the Meta Model API, which TechCrunch notes lands roughly in line with GPT-6 Sol or Opus 5.5. Wonder Full, the stronger of the two, is available today only to Muse Code subscribers and is not sold through the API at all, per Meta's blog post. Both models are slated to power Meta's consumer Muse agent, rolling out first to Muse subscribers.

The flagship stays off the API price list

Developers comparing frontier models on cost now have a Meta entry at parity with GPT-6 Sol and Opus 5.5, but only the lesser of the two new models. Wonder Full, the variant Meta positions as more capable on its own charts, is locked to a Muse Code subscription, so API customers cannot buy their way into the top-tier model the way they can with OpenAI or Anthropic's flagship tiers. That split matters for anyone deciding where to build a memory-dependent agent: a team wanting to test long-horizon recall across sessions can do so cheaply through the API with Wonder, but the model Meta calls its best effort stays walled off inside its own consumer product. Meta also says it trained the Wonder family specifically for discretion with private information and resistance to adversarial instructions, per its own safety section, a direct response to the fact that a memory-persistent agent that acts on a user's messages, calendar, and documents carries more exposure than a stateless chatbot. Every one of these claims (the benchmark rankings, the safety robustness, the capability framing) comes from Meta's own testing; no outside lab has reproduced the MemoryBench or FocusBench numbers yet, so the ranking is a claim to watch rather than a settled result.

Context and reactions

Wonder and Wonder Full are the first frontier-class release out of Meta's AI effort since Scale AI founder Alexandr Wang restructured the group in August 2025, per The New York Times, as cited in TechCrunch's report. In the year since, Meta shipped the lightweight Muse Spark model and the consumer Muse agent but had not put out a model aimed squarely at the state-of-the-art tier the way OpenAI and Anthropic routinely do. Meta's own post dates the effort precisely, saying the restructuring began "fourteen months ago" and that the lab has gone from "rebuilding our stack from first principles to pushing the frontier," per Meta's announcement. The company frames Wonder as a step up that scale, calling the pair "the next step on our scaling ladder and a major leap toward personal superintelligence." Meta also disclosed that the models were trained at its Prometheus data center in central Ohio, which came online earlier this year, per TechCrunch, giving the release a concrete infrastructure tie-in beyond the benchmark claims.

What to watch next

Watch whether any independent lab publishes a MemoryBench, FocusBench, or general-benchmark run on Wonder that confirms or undercuts Meta's numbers, and whether Meta eventually opens Wonder Full to the API the way it broadened access to Muse Spark earlier this year. A gap between Meta's claimed memory lead and third-party replication would be the clearest signal of how much of this launch is substance versus framing.

Sources