Meta introduced Muse Wonder and Muse Wonder Full on 12 October. By its own account these are the company’s first models that compete at the frontier, built for personal and professional agents. Alexandr Wang, who has run the AI group since its overhaul, gave CNBC the codename: Watermelon.
Active Memory and Active Focus
Meta hangs the models on two terms. Active Memory is meant to carry what matters across conversations and pull up the right detail at the right moment. Active Focus describes staying with a task: trying other routes, not stalling at obstacles, asking before anything consequential happens.
That is exactly where agents fall over in daily use, and Wang says as much. “When you do real-world tasks, you’re going to get roadblocks and you’re going to get things that come up or various setbacks,” he told CNBC. The model was trained to try a different approach instead of stalling and checking back in. Meta’s examples are dinner reservations and availability checks.
The measuring sticks are MemoryBench, FocusBench and SWE-Marathon. On Meta’s own numbers, Wonder and Wonder Full match the benchmark scores of Claude Fable 5.1 and GPT-6 Astra, and come out ahead on MemoryBench. Both were trained at the Prometheus datacenter in Ohio, which came online this year. Meta says nothing about context window or parameter count.
Who gets to use it
Muse Wonder runs through the Meta Model API and in Muse Code. Muse Wonder Full is limited to Muse Code subscribers and isn’t served over the API. Inside the Muse app, Wonder reaches subscribers first, and those pay $20 or $100 a month. Muse Code runs $50 or $200.
Through the API, Wonder costs $4.00 per million input tokens and $20.00 per million output tokens. That puts it roughly where GPT-6 Sol and Claude Opus 5.5 sit.
One more line that fits the rest of the week: Meta says Wonder carries its strongest safeguards yet against CBRN and cyber misuse, while keeping frontier cyber capabilities available for defensive work. For now those safeguards only ship with the paid model. A fuller safety report is promised.
Two statements about price that don’t line up
Wang told CNBC the API option would cost more than the standard subscriptions, because it’s a much larger frontier-level model, but would still be priced “meaningfully below” the leading models from OpenAI and Anthropic. The published price list sits roughly where Opus 5.5 and GPT-6 Sol already are. One of those two isn’t right, and it isn’t a detail: anyone building agents decides on exactly that ratio.
The day still means something for Anthropic. I wrote about Claude Haiku 5.5 and its 10 cents per million tokens last week, which is one direction to compete in. Muse Wonder takes the other: a model selling memory and persistence rather than price. If Meta’s numbers hold up, Anthropic now has two rivals on the agent field instead of one.
Sources
- Meta AI Research: Muse Wonder, A Frontier Model For Personal & Professional AI Agents
- CNBC: Meta debuts Muse Wonder AI model in latest test of users’ willingness to pay
- TechCrunch: Meta releases its first frontier model, Muse Wonder
- SiliconANGLE: Meta releases Muse Wonder AI models built to run personal and professional agents