Meta Reveals Muse Wonder, a Much Larger Frontier Model, and Discloses the $100 Billion-Plus Bet Behind It
Zuckerberg and Wang detail the new model, its paid-first rollout and the economics of a gigawatt-scale AI program on a podcast hosted by Stripe's John Collison, October 13, 2026
Meta Platforms is launching Muse Wonder, the first model from Meta Superintelligence Labs built on a new, significantly larger pre-train. CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang discussed the model, the cost of building it and the commercial plan in a long conversation with Stripe co-founder John Collison. The most useful disclosures were not the benchmark claims, which are Meta's own, but the economics of the program and the admission of how little visibility the company had when it committed capital. Collison also noted that Stripe is acquiring OpenRouter, where Meta's models are distributed, and that Stripe worked on the Muse commerce integration. Investors should read his questions with that relationship in mind.
Muse Wonder: a bigger base model, sold first to paying users
Wang described Muse Wonder as "a new class of model that is at the frontier." It comes in two versions, Muse Wonder and a higher-end Muse Wonder Full. The earlier Muse Spark line was built on a pre-train completed at the end of last year. Wonder rests on a separate and "significantly bigger" pre-train, which Meta has been post-training for some time. Internal code names were Avocado for Spark and Watermelon for Wonder. Zuckerberg said the company has "run out of fruits" and will need a new naming convention.
Wang said Wonder is state-of-the-art on Meta's own evaluations of personal and professional agent tasks. He also claimed leading results on DeepSWE and SWE-Marathon, strong results on Terminal-Bench, and top results on several finance and legal evals. Meta highlighted three design priorities. The first is what it calls active memory, built into the model itself rather than a surrounding harness. Wang said memory is where early agent products have been weakest, pointing to the long-term and short-term compaction approach in OpenClaw. "Our evaluations show this, but this is something that we're quite proud of," he said, while acknowledging "there's still plenty of room to improve." The second is active focus, or persistence across long, ambiguous tasks. The third is domain expertise in areas such as finance, law, health and tax, which Wang said should make the agent feel "like a team of experts in your corner."
The commercial rollout is cautious. Because Wonder is larger and more expensive to serve, Zuckerberg said it will start "primarily in paid subscriptions" inside the Muse app, while Muse Spark remains free with what he called a huge allotment of free tokens. Wonder will also be available through the Meta Model API and OpenRouter. Muse Code subscribers will get Wonder Full. Zuckerberg said Meta hopes "not too far from now" to deliver a model as good as Wonder in the Spark form factor, which is the version that could serve hundreds of millions of users. Until then, Meta's best model does not reach its free user base, and the company is still working out how to serve it at scale. Zuckerberg also said Wonder's coding gains will matter to a minority of Muse users.
How Meta says it is escaping benchmark gaming
Asked about Goodhart's Law and benchmaxing, Wang said Meta increasingly steers training toward internal evaluations. External benchmarks, he said, are "not a crazy place to start" but are mainly useful for "looking for signs of life." He was blunt about their provenance: relying on them means "outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark."
Meta's research post quantifies what it calls "big model smell," defined as the gap between a model's performance on tasks resembling its RL and post-training data and its performance on out-of-distribution tasks. Wang said that gap narrows sharply from the Spark line to the Wonder line, which he attributed to the stronger base model. The metric is Meta's own and unaudited. It does suggest the company sees the larger pre-train as the source of the gain, not just more post-training. Wang cited Spark's progression from version 1 to 1.3, which used no new pre-training, as evidence that reinforcement learning recipes are improving quickly across the industry. Meta says it is scaling both pre-training and RL at once.
Zuckerberg framed the product philosophy bluntly: "the model is the product." The hard product work in Muse, he said, is the privacy and security model for the data agents touch, while personality and capabilities are trained in. He pointed to discretion as a trait that matters more for a personal agent than for a coding agent. A personal agent must know a lot about the user while "disclosing as little as possible" to third parties, for example when booking a restaurant around a medical condition. Muse has no model or effort selector, a deliberate choice Zuckerberg tied to confusing naming across the industry.
The $100 billion decision made before the models existed
The most striking disclosure was how the program was financed. Zuckerberg said that at the end of last year, with the models not yet seen, Meta was preparing to write a check of "like, I don't know, like $100 billion, $200 billion." He recalled asking Wang directly: "Alex, are you sure this is going to work?" Wang said that at an executive offsite in Hawaii, CFO Susan Li repeatedly pulled him aside to say that "if this doesn't work... it's going to be very, very bad in financial terms." He described the situation as "entrepreneurship on hard mode. It's the opposite of Lean Startup," because waiting for proof means being "radically behind."
On cost, Zuckerberg gave rough industry-style figures. Operating a 1 gigawatt cluster for a year runs on the order of $15 billion, and building one costs roughly $50 billion to $60 billion upfront. Research costs, including testing at every rung of the scaling ladder, are probably comparable to the cost of the final run. Serving costs come on top. He said the financial roll-ups he has seen make "tens of billions of dollars" a good estimate for training a frontier model. Meta's Prometheus cluster in Ohio is scaling beyond 1 gigawatt, and the Hyperion cluster in Louisiana is planned to reach 5 gigawatts. Zuckerberg said both were planned in 2023 or 2024, before MSL existed. He added that data center capacity for serving Muse had to be committed years ahead of a product that did not yet exist.
Zuckerberg argued the downside is smaller than the numbers suggest. If Meta's models fell short, he said, other buyers would pay for the capacity at roughly that gigawatt price, so "AI overall has to fail" for the compute to be stranded. That is a management view, not a guarantee, and it depends on continued industry demand for compute.
How Meta says it will pay for it
Asked how the buildout gets funded, Zuckerberg named four revenue paths. First, subscriptions, starting with Wonder access. Second, commerce, where Meta would take "a small cut" of what Muse saves or earns users, along with a "business in a box" agent for small businesses. Third, the API and a new Meta Enterprise Platform led by CJ Desai, aimed at enterprises Zuckerberg expects to want an "excellent and affordable" Wonder. Fourth, the core advertising business, which he said is "hundreds of billions of dollars and still growing extremely quickly" and can be accelerated "a few percent faster" with these models. No revenue figures or timelines were given for any of these.
He also pointed to an accounting advantage over cloud-reliant rivals. Meta capitalizes and depreciates its own data centers, whereas researcher salaries are expensed. He said even H100 GPUs are more valuable now than last year because compute remains constrained. Meta's stated model is to build to a billion users first and scale revenue afterward, and the answer to whether AI will make money "better be yes."
Agentic commerce: early behavior, not yet a financial line
Zuckerberg said Muse is one of the first products in which users fully delegate purchases to an agent. Meta trained browser use directly into the models so no merchant has to integrate, and it is pursuing direct API access with major partners. He said "way more businesses" are calling to integrate deeply than are trying to block agents. His personal example was a weekly baking routine with his daughter, with Muse ordering ingredients through Instacart. He said he opened an Instacart account solely for Muse and now uses the service once or twice a week. His argument was that agents with their own accounts could expand the total addressable market of commerce, not just shift share. This is anecdote, and Meta offered no data on order volumes or conversion.
Third-party reports, Zuckerberg said, show Muse as possibly the fastest-growing agent or AI app to date, though no user figures were provided. He added that Meta did not have growth tooling ready at launch because it expected to iterate toward product-market fit.
A small, talent-dense lab, and an admission about Llama 4
Zuckerberg was direct about the reset. Meta launched Llama 4 on April 4, 2025, was dissatisfied with the trajectory, and began calling outside advisers about 2 weeks later. Wang said the research team is "less than 200 people," probably closer to 100 excluding infrastructure staff, versus hundreds elsewhere, with compute concentrated on a few bets and not "peanut buttered" across competing efforts. He described one early meeting in which a Llama-era lead asked whether Meta even needed a reasoning model, calling it a "heart-sinking moment." Zuckerberg said the lab had to feel like a restart in order to recruit the best people. He also said he cleared his calendar for a month or 2 to meet researchers. Wang noted that Wonder was planned from August and that early post-training gains came quickly, which he called a good sign.
Why Meta argues AI will not be winner-take-all
Wang said the recursive self-improvement runaway predicted in 2023 has not happened. He attributed this to uneven execution at every lab and to distillation, which he called an "anti-hegemonic force" that packs frontier capability into smaller models. Zuckerberg said Meta cares more about personal and professional agents than other labs, which he described as differentiation. He added that a world with diverse models and checks and balances is safer than one where a few companies control the best models. Zuckerberg also said Meta may not build an app store for its glasses, which he expects to be a "pure multimodal product," and that Muse Gadgets has drawn interest from phone makers.