How Meta’s Muse Wonder Plans To Shake Up The AI Agent Market

For the past year, the artificial intelligence conversation has largely revolved around who can write the best code or generate the most convincing image. Meta is attempting to pivot that narrative away from static answers and toward sustained action. With the launch of the Meta Muse Wonder and its higher-tier sibling, Muse Wonder Full, the company is betting that the next era of AI isn’t about answering a single prompt perfectly, but about carrying a complex, multi-step project across days or weeks without losing the plot.

Developed internally under the codename “Watermelon,” the models represent a massive overhaul of Meta’s AI infrastructure. Spearheaded by CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang—who joined following a massive acquisition deal to reshape Meta’s AI division—this release is the culmination of fourteen months of restructuring. The goal wasn’t just to build a smarter chatbot, but to create an engine capable of powering “Muse,” the company’s ambitious personal agent platform.

Remembering You When You Forget

The most glaring flaw in current AI assistants is their amnesia. You explain your complex legal situation on a Tuesday, and by Thursday, the assistant has completely forgotten the context. Muse Wonder attempts to solve this through what Meta calls “Active Memory.” Instead of treating every chat window as a blank slate, the model actively retrieves relevant details from past interactions and carries them forward.

Combined with “Active Focus,” the system is designed to handle the friction of real-world tasks. If you ask it to navigate a convoluted insurance claim or analyze a dense commercial lease, the model doesn’t just spit out a summary and wait for your next command. It attempts to push through blockers, tries alternative approaches when it hits a wall, and even asks for clarification before executing a sensitive action.

During early demonstrations, Alexandr Wang revealed his own personal agent—affectionately named “Oiler”—routinely sources obscure teas from small farms in Japan and Taiwan. Meanwhile, Zuckerberg reportedly uses his instance to coach his mixed martial arts training by ingesting raw gym camera footage and writing its own video-processing pipelines to analyze his form. The pitch here is that the model itself is the product, prioritizing proactivity and personality over endless menus of user settings.

The Price Of Thinking Harder

While the capabilities are impressive, independent benchmarking reveals a mixed bag. According to data from Artificial Analysis, Muse Wonder Full scores exceptionally high on agentic evaluations, jumping to 56 points on their Intelligence Index and doubling its score on Terminal-Bench for autonomous problem solving. In coding arenas, it ties with top-tier competitors like Anthropic’s Claude.

Meta’s engineering teams spent months training the models on long-running software projects to ensure they wouldn’t abandon a task halfway through just because the codebase became too complex. This sustained execution capability is precisely what independent testers look for when evaluating whether an AI can genuinely replace junior developers or administrative staff, rather than just serving as a glorified search engine.

However, raw factual knowledge remains a weak spot. On the AA-Omniscience test, which measures encyclopedic recall, the Wonder models trail behind the latest offerings from OpenAI and Google. To compensate, Meta has heavily optimized the model to hallucinate less frequently, boasting one of the lowest error rates in its class.

This leap in performance comes with a steep premium. Running a more persistent, context-aware agent requires significantly more computational power. API pricing for Muse Wonder has climbed to nearly three times the cost of its predecessor, Muse Spark, sitting at $4 per million input tokens and $20 for output. For developers and enterprise users building complex software on top of the Meta Model API, that cost is justified by the reduction in manual oversight. For the average consumer, it means these frontier capabilities will likely remain gated behind the paid subscription tier of the Muse app.

The Infrastructure Bet

Powering an AI that can process live video feeds and remember your life’s context requires an astronomical amount of electricity and silicon. Meta has been quietly constructing the “Prometheus” gigawatt-scale data center in Louisiana to handle this exact workload. Running a single gigawatt cluster costs roughly $15 billion annually, and the company has signaled a commitment of up to $200 billion in capital expenditure to ensure they aren’t left behind by rivals.

The shift from experimental research to reliable consumer utility requires moving beyond temporary cloud rentals into dedicated, physical fortresses of computation. This physical footprint ensures that when you ask your personal agent to untangle a decade of messy financial records, the system doesn’t time out or throttle your request during peak evening hours.

Wang has been blunt about the stakes, noting that for this specific bet to fail, the entire premise of artificial intelligence would have to fail first. The industry is no longer just competing on who has the smartest algorithm, but on who can sustain the most massive, uninterrupted physical infrastructure to keep it running.

You open a new tab to pay a frustrating utility bill, fully expecting to spend the next twenty minutes navigating a maze of automated phone trees and broken web portals, only to find your browser has already negotiated the payment, filed the receipt in your tax folder, and scheduled a reminder for next month without you ever typing a single word.