Summary: Meta launched Muse Wonder and Muse Wonder Full on October 12, 2026, and pitched them as frontier models tuned for personal and professional agent work rather than just raw benchmark scores. The bet: agents win on memory, sustained focus, and domain expertise, not one-shot answers. For builders, the catch is that only the standard Wonder is API-accessible, so what you can actually ship differs from what Meta demos.
What actually shipped
Meta Superintelligence Labs built the Wonder series, training it on the gigawatt-scale Prometheus data center and pulling together research from multiple internal teams, according to TechOrange’s report. The Financial Times framed it as Meta’s return to frontier competition after heavy spending on AI talent, data centers, and compute.
Notably, Meta moved away from its Llama-era open-weights playbook. Wonder is a closed, commercial model, and standard Wonder’s API pricing — $4 per million input tokens and $20 per million output tokens per VentureBeat — matches Claude Opus 5.5. Meta is not competing on price.
On Artificial Analysis’ composite intelligence index, Wonder Full scored 56, beating GPT-6 Astra’s 53 but trailing Claude Opus 5.5’s 58. Competitive at the frontier, not dominant everywhere.
The agent-first pitch
What Meta emphasized instead is a set of agent capabilities: domain expertise, Active Memory, Active Focus, and frontier coding.
- Domain expertise: On the Harvey LAB-AA legal task benchmark, Wonder Full scored 13.1 versus GPT-6 Astra’s 8.6 and Claude Opus 5.5’s 4.2.
- Active Memory: The model retains context across conversations — preferences, constraints — and retrieves it when a new task needs it, so users stop re-explaining themselves.
- Active Focus: When blocked, the model tries alternatives, revises plans, and checks in with the user. Wonder Full hit 54.2% on FocusBench, ahead of Astra (47.7%) and Opus 5.5 (45.8%). Standard Wonder managed only 44.9%, a real gap between the tiers.
- Coding endurance: On SWE-Marathon, Wonder Full scored 57.5 and Wonder 53.1, both above Opus 5.5’s 46.3.
Meta also stressed generalization — handling unfamiliar task combinations it wasn’t specifically trained on. For personal agents, where each user’s goals, tools, and constraints are unpredictable, that matters more than acing known test items. Meta says Wonder Full beats what trend extrapolation from Muse Spark 1.3 would predict.
I covered the memory-and-focus framing in more depth in Meta’s new agent model treats memory and focus as first-class features — worth a read if you’re weighing what these capabilities mean for agent architecture.
Why Meta needs this to work in a consumer product
The scores feed directly into Muse, Meta’s personal agent, which the Wall Street Journal reports has roughly 9 million downloads but slowing growth and retention behind ChatGPT. Meta says Wonder will first strengthen complex workflows and deliverable generation for subscribers, with demos like planning a running route across browsers and apps based on user preferences.
That’s the real test. A model that scores well on FocusBench only matters if it makes Muse an agent people keep using week after week.
What you can actually integrate today
Here’s the constraint for anyone building on this: Wonder Full is available only to Muse Code subscribers and is not on the general API, per VentureBeat. What enterprises can integrate is standard Wonder — the weaker tier, and notably the one that underperforms competitors on FocusBench.
Active Memory also doesn’t mean you get enterprise-grade long-term memory for free. If your agent needs to pull context across systems, documents, and personal data, you still own the data-access permissions, tool integrations, and security work.
The practical takeaway: benchmark the standard Wonder API against your own long-horizon tasks before assuming the Full tier’s numbers apply to what you’d ship. And if you’re routing agent traffic through a gateway, Meta’s agent model is already reachable through Vercel AI Gateway, so a side-by-side comparison is cheap to run.
Whether Meta’s agent claims hold up in real enterprise environments — messy permissions, flaky tools, ambiguous instructions — remains to be seen. The direction is clear, though: the frontier race is shifting from who answers best to who sustains work longest.