Meta’s Superintelligence Lab has officially escalated the AI arms race. On Monday, the company released Muse Wonder and Muse Wonder Full, marking its first true foray into frontier models. This is the first major, state-of-the-art release since Scale AI founder Alexandr Wang restructured Meta's AI division in August of last year, signaling a shift in how Meta plans to compete directly with industry leaders.
Until now, Meta’s recent footprint in the space has been defined by Muse Spark—a lightweight model optimized for low-latency tasks—and its consumer-facing Muse agent. The Wonder models pivot in the opposite direction. Trained at Meta’s newly operational Prometheus data center in central Ohio, these models are larger, built for complex tasks like coding, and command higher benchmark scores. According to Meta's model card, Wonder and Wonder Full are achieving parity with current top-tier models, posting benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
Pricing and API Access
For businesses building custom operational workflows, access and cost are the immediate hurdles. Muse Wonder is available right now through the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens. This cost structure places it firmly in the premium tier of AI models, making it roughly equivalent to running GPT-6 Sol or Opus 5.5.
Interestingly, Meta is keeping its most powerful system close to the chest. Muse Wonder Full will not be served through the API for custom development. Instead, it will be strictly available to Muse Code subscribers. While both models will power Meta’s own Muse agent, businesses looking to build their own proprietary automations will have to rely on the standard Wonder model.
The Active Memory Advantage
The most compelling technical detail of the Muse Wonder release is not just raw reasoning or coding capability—it is how the model manages context over time. Meta has placed a massive emphasis on what it calls "active memory," designing the models to retain and synthesize user information over extended periods.
To prove this, Meta highlighted the MemoryBench benchmark, a framework specifically built to challenge continuous learning and memory construction in AI models. According to the company's testing, Muse Wonder currently scores higher on MemoryBench than Fable, Opus, or Astra. The goal, as research scientist Kavya Kopparapu explained in the launch video, is to ensure the model can "remember what’s important to you over days, weeks, months, and beyond."
What Muse Wonder Means for SMB Automation
For small and mid-sized businesses, a new top-tier model from a major player is always a net positive—more competition inevitably drives down API costs across the board. But the specific capabilities of Muse Wonder offer distinct operational advantages.
First, the improved performance on complex tasks like coding directly translates to better operational automation. When our team builds internal workflows for SMBs, the AI often needs to act as a translator between different software platforms. A model that excels at coding is typically much better at consistently generating correctly formatted structured data (like JSON) and reliably triggering external APIs without hallucinating formatting errors.
Second, and most importantly, the focus on "active memory" could fundamentally change how persistent agents are built. Currently, if an SMB wants an AI to remember a client's specific preferences across a dozen different support tickets, or track an ongoing vendor negotiation over the course of a quarter, developers have to build complex workarounds. We have to set up external databases, build retrieval-augmented generation (RAG) pipelines, and manually feed the AI its own historical context every time it triggers.
If Muse Wonder’s active memory works as advertised in production, it drastically simplifies this architecture. A customized HR assistant could natively remember a new hire's onboarding progress over a 30-day period. A vendor management bot could pick up a conversation exactly where it left off a week ago, synthesizing past emails without requiring expensive context-window stuffing.
At $4 per million input tokens, Muse Wonder is far too expensive to be used for simple, high-volume inbox routing or basic data extraction. Those tasks still belong to smaller, cheaper models like Muse Spark. But for complex, long-running workflows where historical context and reliable data structuring are critical, Meta has just provided a highly competitive new tool for the automation stack.