Meta's newest AI model is designed to remember you. Muse Wonder, announced on October 12, 2026, is built around what Meta calls Active Memory: it keeps important context across conversations and brings back relevant details later. The hosts of the Moonshots podcast, recording the next day, disagreed about whether the model was a real return to the front of the AI race. They also asked a second question: does it matter more that an assistant is smart, or that it knows everything about you?
Host Peter Diamandis, founder of XPRIZE and Singularity University, opened with a post by Garry Tan, the CEO of startup accelerator Y Combinator, who wrote that his bet was that Muse would win. Diamandis gave two likely reasons. The first was money: he said Meta was on track for $250 billion in revenue and more than $80 billion in net income in 2026, and was spending it on computing power. The second was reach. According to Diamandis, 3.6 billion people use Meta's apps at least once a day, "nearly half the planet."
What Meta launched
Meta released two versions, Muse Wonder and Muse Wonder Full. Alongside Active Memory, the company describes a feature called Active Focus. It is meant to keep the model working when it hits a blocker: it tries alternative approaches, adapts as information changes, and asks for clarification or confirmation before important actions. Meta's demonstrations include planning a run, organizing wedding seating and building software. The company pitches the models for personal tasks and for professional work such as leases, insurance, financial information and long-running software projects, according to its launch post.
The model is not yet in Meta's consumer app. Wonder is available to developers through the Meta Model API, a paid interface that lets software call the model, and through Muse Code, Meta's coding product. Wonder Full is available to Muse Code subscribers. Meta says Wonder will power experiences in its consumer Muse app later, starting with subscribers.
How Meta got here
Diamandis gave the backstory. In April 2025, Meta's Llama 4 model got what he called "a very lukewarm reaction," and Meta looked as if it had fallen behind. In June 2025, he said, Mark Zuckerberg spent $14.3 billion to buy 49% of Scale AI, a company that supplies data used to train AI models. Zuckerberg also brought in Scale's 28-year-old founder, Alexandr Wang, as chief AI officer of the new Meta Superintelligence Labs. A talent war followed, Diamandis said, with Meta and OpenAI both offering some researchers $100 million signing bonuses. Critics called the deal "the most expensive acqui-hire in history," he said, meaning a purchase made mainly to get the people.
Scale's own announcement from June 12, 2025 describes the deal differently. It says Meta's investment valued Scale at more than $29 billion and made Meta a minority shareholder. Wang moved to Meta but stayed on Scale's board, and Scale said it would keep operating independently and serving other AI labs, businesses and governments.
Blundin: the expensive bet paid off
Asked whether the deal was paying off, Dave Blundin, founder and general partner of Link Ventures, said: "Oh, God, yeah. Time and money well spent." He told a story to back that up. By his account, less than a week after the hosts met OpenAI's Mark Chen, Chen got a call from Zuckerberg offering $1 billion to leave OpenAI for Meta, and he said no. Blundin said Chen had been the person who first persuaded Zuckerberg that, with Meta spending so much on computing, "a couple of very expensive people will make all the difference." Blundin said that is how it turned out: "The price seemed utterly insane. It's not even vaguely insane."
Blundin also contrasted the new team with Meta's earlier AI leadership, including Yann LeCun, whom he described as more interested in breakthroughs and his reputation as an inventor. As Blundin sees it, building a model like this is "really just an engineering problem," a very complicated one, and Meta finally hired people with an engineering mindset. His advice to other companies that want to stay relevant: "Just do what Mark did."
Wissner-Gross: below the frontier once cost counts
Alexander Wissner-Gross, a computer scientist and founder of Reified, was much less impressed. When Diamandis put up Meta's benchmark chart and asked what impressed him, he answered: "Almost nothing."
His complaint was about cost. AI companies call their strongest models "frontier" models. Wissner-Gross argued that high scores on capability tests are easy if you put enough computing power behind each task and ignore how many dollars, tokens or seconds it takes. Tokens are the small chunks of text a model reads and writes, and providers charge by the token. Ignore cost, he said, and "you can just put monkeys at typewriters and eventually get the right solution." In raw capability, he put Muse Wonder roughly level with Anthropic's recent Sonnet 5.5 or Opus 5.5. "But if you pay attention to cost, actually, it's below the frontier," he said.
He thinks Meta badly needs to claim a place at the frontier. In his view, Meta's long-term strategy rested on two bets, AI and augmented and virtual reality. The second "hasn't panned out," and investors have punished Meta for that. "That leaves AI," he said. He wants Meta to focus on capability per unit of cost, and he said he would like the frontier to be truly competitive.
An independent evaluation by Artificial Analysis, published on launch day, shows both sides of the argument. On its Intelligence Index, Wonder scored 52 and Wonder Full 56 at maximum reasoning effort, up from 48 for Meta's earlier Muse Spark 1.3. On Terminal-Bench 4.0, a test of agents doing real work at the command line, the text-based interface programmers use to control a computer, the two new models succeeded 52% and 57% of the time, compared with 33% for Spark. The index is not a single exam. It is a weighted blend of ten mostly English, text-based evaluations: 30% agent tasks, 20% coding, 20% scientific reasoning and 30% general capabilities, according to the evaluator's methodology.
The cost figures partly support Wissner-Gross. Wonder cost $4.52 per weighted index task, against $1.60 for Spark and $0.72 for OpenAI's GPT-6.1 Sol, more than six times as much as the OpenAI model. It was still cheaper than Claude Opus 5.5, at $5.98, in the configurations tested. About 76% of Wonder's cost came from the large amount of input it reads. Wonder Full has no per-token price, so the evaluator could not give a comparable cost for it. The evaluation also identified factual knowledge as a relative weakness of the models, while noting comparatively low hallucination rates, meaning fewer confident but made-up answers.
Where is the revenue?
Wissner-Gross saw a deeper tension in Meta's strategy. He said Meta seems to be "speedrunning the lesson that OpenAI had to learn the hard way": the revenue comes from businesses, not consumers. While OpenAI tries to become more like Anthropic, "the enterprise AI company," Meta is moving into the consumer gap. He pointed to how heavily Instagram advertises the Muse app to anyone who opens it. "Where's the revenue? I think that that's an elephant in this room," he said.
He offered what he called the steel-man case, the strongest version of the opposing argument. Meta may need to show huge uptake, perhaps a billion monthly users, to persuade shareholders to keep paying for data centers and stronger models. But he doubts those users will turn to Muse for hard math or programming. He expects them to use Anthropic, OpenAI and Chinese models for demanding reasoning, and Muse for what he called "personal sycophancy," a chatbot that flatters and agrees with its user, which he called "probably a killer app" for personal chatbots.
That raises a question about the Scale deal, he said. In his view, Meta bought into Scale "entirely for the post-training data." Post-training is the stage after a model's initial training where it is refined for particular skills and behavior. That work mostly pays off in high-end business uses, he argued, not in the companionship and flattery he said have been reported as popular uses. "For Muse, what's the point? I think they're going to have to sort this out."
Ismail: distribution plus memory
Salim Ismail, founder of Open ExO, sided with Tan: Muse "could" win. Meta has long experience delivering easy-to-use products to huge numbers of people, he said. The company's comeback also shows that "today's AI leadership is not a permanent moat": with the right capital, talent and infrastructure, a company can change the landscape. People once doubted that social networking could support an advertising business, he said, but Facebook knew so much about users' social graphs, their networks of connections, that it did.
He thinks memory is the new version of that advantage. A model that remembers "your preferences, your goals, your relationships, your ongoing projects" would be "massively more useful than a slightly smarter model" that has to start every conversation from scratch, he said. He described the change as a move from chatbots to "a personal operating system," which brings trust problems. Facebook, he said, has "deliberately and steadily eroded the trust" over the years. "They could actually win and the privacy people are going to freak out."
The lock-in problem
Diamandis asked how people should choose among the crowd of personal agents, naming OpenClaw, Hermes, GrokBot, Muse and Dots. Blundin said to decide soon. Because an assistant like Muse keeps accumulating your preferences, "all your personal data is piling up inside," and switching gets harder over time. "Right now, they don't make the data portable," he said. Choosing to stay unlocked, by keeping your data portable and using only open-source models, is "a perfectly reasonable choice, too." But he pointed to the launch demonstration of the assistant filling out an auto insurance form. That is convenient, he said, only because it knows all your data, and he called that "the addictive nature" Zuckerberg is going after.
Diamandis then raised a Wall Street Journal article titled "Personal AI agents are great until they share your bank statement in the work chat," describing it as a warning about agents that overshare, overpromise and buy things you did not intend. He cited stories of Muse making an unwanted purchase. Katherine Bindley's reporting, republished by Mint, gives a more specific picture. The unapproved purchase, a swimsuit bought with store credit that bypassed the user's credit-card approval step, was made by an Instinct agent, not Muse. The Muse story came from Shalini Dinesh. She found it useful for organizing a birthday party until an RSVP error merged two children who shared a first name. In another case, Shane Mac said a Grok Bot agent posted personal financial information to his company's Slack after confusing similarly named groups; he later said its team had fixed the problem. These are individual experiences, not measured failure rates.
Ismail called the article "a little exaggerated" but said it pointed to something larger. Social networks already lock people in: try getting your connections out of LinkedIn, he said, "without paying a fortune." Now that social-graph lock-in is being replaced by "a much more intimate one," because "your personal memory belongs to an AI company." In his view, "the next trillion dollar battle is not going to be about which AI is going to be smartest. It's which one knows you best."
He offered a tip. For the past few months, after each major set of queries to a model, he has run a separate request asking it to pull out the key concepts, relationships and entities and turn them into "a personal ontology," a structured personal memory. Because he keeps it himself, he can carry it from one model to another and avoid lock-in.
Some agents already treat memory as something the user can see and edit. Nous Research's Hermes Agent, one of the agents Diamandis named, stores its persistent memory in two small files: one for project and environment knowledge, capped at 2,200 characters, and one for user preferences, capped at 1,375. A snapshot is loaded at the start of each session. When a file fills up, the agent has to consolidate or remove entries rather than silently drop them.
One agent, or many?
Diamandis asked whether a single AI might eventually sit on top of all the others and pick GrokBot, Muse or Dots as needed. Wissner-Gross doubted it. He called these "pretty contrived artificial layers" that will "dissolve with time": the companies want to be vertically integrated, owning every layer, and "the operating system wants to become AI."
Later in the episode, when the conversation turned to whether one system could plug into whatever AI is best, Ismail made a different prediction. Today, he said, every platform is competing to put "the most cute little cuddly face" on its agent because it needs to own the context layer, the store of what the assistant knows about you. He expects people to move instead to an agent they own themselves, probably an open-source one built on open standards so it can move between models, with no need for the user to know whether Claude, GPT or Grok handles a given task. "So you're going to end up with one trusted agent with access to the world's intelligence."