Meta has released Muse Wonder and Muse Wonder Full, two new models from Meta Superintelligence Labs that the company describes as its most capable yet. TechCrunch calls the pair Meta's first real attempt at a frontier model, the tier of AI that OpenAI, Anthropic and Google have been trading blows in for the past few years. The launch was announced on October 12, and unlike many big model reveals, at least one of the two is something developers can start using right away.
The pitch is not "a smarter chatbot." Meta is framing Muse Wonder around personal and professional agents: software that remembers what matters to you, keeps working on a task for a long time, and takes actions across apps on your behalf.
What Meta actually released
- Muse Wonder is available now through the Meta Model API and Meta's coding product, Muse Code. According to TechCrunch, API pricing is $4 per million input tokens and $20 per million output tokens, which puts it in the same price band as the top models from OpenAI and Anthropic.
- Muse Wonder Full, the larger variant, is available only to Muse Code subscribers. TechCrunch reports it will not be offered through the API.
- Both models will power Meta's Muse agent, with subscribers getting access first. Meta also says it plans to use Muse Wonder to build smaller specialized models for the wider Muse audience later.
Meta says the models were trained at its gigawatt-scale Prometheus data center, which TechCrunch places in central Ohio. In its announcement, Meta notes that it began a "ground-up overhaul" of its AI efforts fourteen months ago. Until now, the visible results of that rebuild were the lightweight Muse Spark model and the Muse agent. Muse Wonder is the first release aimed squarely at the top of the leaderboard.
The two ideas Meta is leaning on: Active Memory and Active Focus
Most of Meta's write-up is about two capabilities it has given names to.
Active Memory is the model's ability to carry important context about a person across conversations and pull up the right detail when it becomes relevant later. That is a direct answer to one of the most common complaints about AI assistants: you explain your situation once, and next week you have to explain it again. Per TechCrunch, Meta's own testing shows Muse Wonder scoring higher on the MemoryBench benchmark than the leading models it compared against.
Active Focus is about persistence. Meta describes it as the model continuing to work toward a hard goal by trying alternative approaches, working through blockers and adapting when new information shows up, while still asking for clarification and checking in before important actions. Meta says it measures this with FocusBench and with SWE-Marathon, a test of sustained work on long-running software projects.
Meta also highlights domain expertise in areas like law, finance and health, citing Harvey's Legal Agent Benchmark, which asks models to work from source documents and produce legal work products. The demos Meta showed are telling in their ordinariness: planning a Saturday run around someone's preferences, putting together a wedding seating chart with complicated constraints, and walking through an insurance claim. These are chores, not science fiction, and that is the point.
How it stacks up
Meta claims state-of-the-art results on benchmarks for personal and professional agents, memory and coding. TechCrunch, reading Meta's model card, puts it a little more conservatively: Wonder and Wonder Full achieve comparable scores to the current top models from competitors rather than clearly beating them across the board. As always with launch-day numbers, these are the company's own evaluations, and independent testing over the coming weeks will matter more than the charts in the announcement.
One part of Meta's post is worth a closer look. The company directly addresses "benchmaxxing," the suspicion that models are tuned to score well on popular tests without being broadly capable. Meta says it compared performance on tasks it targeted during post-training with tasks it did not, and found that Muse Wonder Full does better on the unfamiliar tasks than its earlier Muse Spark trend would predict. Meta credits a larger, stronger base model for that. It is an interesting claim, and also exactly the kind that outside researchers will want to probe.
Safety and privacy
An agent that reads your messages, calendar and documents raises obvious questions. Meta says it trained the Muse Wonder family to resist adversarial instructions and handle private information carefully, and that the model has its strongest safeguards to date against chemical, biological, radiological, nuclear and cyber misuse, while keeping cyber capabilities available for defensive use. A fuller safety report is promised "soon," so for now those claims rest on Meta's description alone.
Why it matters
For the past year, the frontier race has largely been a three-way contest. Meta, with its huge compute budget and billions of users, has been notably absent from the very top tier. Muse Wonder changes that, at least on paper, and adds another serious option at the premium price point.
The bigger signal is the direction. Meta is not selling raw intelligence; it is selling memory and follow-through. If the industry's next battleground is agents that know you and finish what they start, then the questions shift from "which model is smartest?" to "which model can I trust with my context, and how well does it keep track of it over months?"
What you can do now
- Developers: try Muse Wonder through the Meta Model API on a real, multi-step task from your own work, not a toy prompt. Compare cost and completion time against the model you already use.
- Teams evaluating agents: test memory deliberately. Give the model context early, come back later, and see whether it recalls the right details without being reminded.
- Everyone else: watch for Muse Wonder arriving in the Muse agent, and pay attention to what permissions it asks for before you connect personal accounts.
Treat the benchmark claims as a starting point, wait for the safety report and independent reviews, and judge it on the tasks you actually care about.
Source: Meta, "Muse Wonder: A Frontier Model For Personal & Professional AI Agents" (Oct 12, 2026); additional reporting: TechCrunch
댓글
댓글 쓰기