Meta’s Muse Wonder Enters the AI Race Carrying OpenAI and Anthropic’s Lunch

Meta’s new agent models promise memory, persistence, and actual follow-through. OpenAI and Anthropic might want to watch out: Zuckerberg has discovered finishing the assignment.

Meta has spent years helping people avoid the thing they were supposed to be doing. Now it has built an AI that promises to keep doing the thing until it is finished. Somewhere inside Facebook, two product managers are fighting over whether this counts as engagement.

On October 12, Meta introduced Muse Wonder and Muse Wonder Full, its most capable models yet, with claimed frontier performance in coding and personal and professional agent work. The sales pitch includes insurance claims, leases, and other tasks previously handled by placing the document on a kitchen counter until it became a structural component of the house.

Wonder launches through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. Wonder-powered experiences in the Muse agent are coming, with subscribers first. Today’s model access and the coming consumer rollout are separate things. The future remains available in several confusingly labeled packages.

My read: OpenAI and Anthropic might want to watch out. Meta is making a serious bid for the valuable part of AI: the moment someone stops admiring an answer and starts trusting a system with a job.

The Llama Has Returned With a Briefcase

Our earlier Muse Spark coverage caught Meta in the awkward process of changing its AI identity while the old identity was still wearing a conference badge. Wonder advances that story from promising reorganization to a direct challenge for serious work.

Meta’s research announcement says the models were trained at its gigawatt-scale Prometheus data center. It also demonstrates a wedding seating chart. Humanity has built an industrial installation named after the titan who gave us fire so that Aunt Linda can sit beyond conversational range of her second husband.

I respect the allocation of resources. Some problems deserve a gigawatt.

The useful context from our guide to the AI model landscape is that a model, a coding tool, and a personal agent are different products. Wonder supplies intelligence. Muse Code gives it a coding workplace. Muse is the personal-agent destination. A naming committee has ensured you will explain this incorrectly at least once.

Active Memory Knows Where You Buried the Task

Meta calls the two central behaviors Active Memory and Active Focus. Memory retrieves relevant context across conversations; Focus keeps a difficult task moving through retries and obstacles, with clarification and checks before important actions.

These sound like qualifications for an office administrator. That is why they matter.

The miserable part of many AI workflows is reintroducing your entire situation to a machine that greeted you yesterday like a lifelong collaborator. You explain the project, the budget, the deadline, and the thing it must absolutely never do. Tomorrow it welcomes you back with the institutional continuity of a goldfish entering witness protection.

If Wonder reduces that burden, it could earn loyalty without winning every abstract intelligence contest. Our personal AI deep dive explored the underlying bargain: remembering you makes an assistant useful and makes leaving it more work. Exporting a chat is easy compared with teaching a replacement why the phrase “quick sync” means the project is in danger.

Persistence matters just as much. A system that confidently proposes eight steps and performs one has invented management consulting. A system that follows through might actually be worth paying.

OpenAI and Anthropic, Please Check the Side Entrance

SiliconSnark has covered OpenAI’s Astra launch and Anthropic’s Fable 5.1 showdown. The rivalry has enough lore to require a streaming-service recap. Previously on Extremely Expensive Autocomplete: everyone became the most intelligent system in the world, briefly.

Wonder makes the competitive question more uncomfortable. What happens if Meta gets sufficiently good at the work people value, then makes that help convenient to reach?

As our Perplexity Computer coverage argued, the race extends beyond the two loudest labs. Customers have the distressing habit of caring about their own tasks more than your corporate rivalry.

Meta already has consumer surfaces where people spend time. Its own Facebook AI guide describes access through the feed, Messenger, and search. Those are existing AI entry points, not a claim that Wonder has rolled out to all of them today. But they illustrate the distribution opportunity: Meta can introduce assistance in places where the user has already arrived to argue about a driveway.

Our Meta Connect roundup traced the broader push into agents, devices, and glasses. A useful model behind those ambitions would give the hardware pitch considerably more substance than “your face could use a subscription.”

OpenAI and Anthropic have their own products, relationships, and reasons users stay. This is a warning about competitive pressure, not an obituary typed during a launch video. Still, treating Meta as a permanently remedial AI student looks increasingly risky. The student has built a power station behind the gym.

The Invoice Has Entered the Chat

According to Meta’s published pricing, Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. Meta lists no long-context premium. These are API prices, not a subscription quote for Wonder Full.

The same documentation says Standard-mode content is not used to train Meta models. Muse Code subscriptions default to Contributor mode, which can be switched off; Wonder’s API has no Contributor option. Apparently the fine print has also acquired Active Focus.

Price matters, but our guide to choosing an expensive robot coworker made the more useful point: count the human cleanup. A cheap answer that needs forty minutes of repair is a luxury purchase disguised as a bargain.

Meta’s AI spending saga and the Hyperion infrastructure story supply the financial backdrop. Wonder is a chance to turn the expensive nouns into useful verbs. Eventually, somebody has to leave the data-center tour and finish a spreadsheet.

Show Me the Work After the Demo Ends

Meta’s research post discusses FocusBench and SWE-Marathon, including sustained software work. It also argues that stronger base models help Wonder Full handle unfamiliar tasks better than Spark’s performance trend would suggest. That is a company-reported case for broader competence. SiliconSnark has not run a controlled Wonder-versus-Astra-versus-Claude test.

The distinction matters because a model can look spectacular under carefully chosen conditions and become a haunted stapler when asked to modify the file you actually have. Our account of AI struggling with PowerPoint remains a useful antidote to the belief that civilization has been solved because a chart has tall bars.

Watch what happens when the request changes halfway through, a tool fails, or yesterday’s remembered preference conflicts with today’s instruction. Does the agent notice? Does it recover? Does it stop before confidently expanding a small mistake into a regional initiative?

Our Grok Bot analysis reached the same practical territory: the tools, permissions, and recovery behavior around a model shape the experience. Intelligence needs somewhere sensible to work. Giving a genius a broken browser is still giving someone a broken browser.

Meta says additional safety evaluations are coming. The useful test will be whether discretion survives the messy jobs memory and persistence make possible. An agent that remembers everything and never gives up sounds wonderful until it misunderstands which wedding you wanted canceled.

Our earlier look at Muse’s Sentinel permission system made the point: the thing deciding what an agent may touch deserves as much attention as the thing generating its ideas. Especially near the guest list.

The Lunch Is Still on the Table

Wonder’s strongest pitch is almost offensively ordinary: remember what matters, keep going, finish something useful. After years of grand declarations about transforming humanity, the industry has discovered that humanity has seventeen unresolved emails and would like to go outside.

That is a market worth fighting over. If Meta can deliver dependable follow-through, it can pressure the companies currently charging for it. The remaining questions are reliability, trust, and whether the product works as well on your annoying Tuesday as it does in Meta’s carefully prepared demonstration.

OpenAI and Anthropic should watch out. Zuckerberg has arrived at the frontier carrying a clipboard.

He says he is here to help with your paperwork. I would label the lunch in the refrigerator.