Meta Releases First Frontier Models: Muse Wonder and Wonder Full

Meta shipped two frontier models, Muse Wonder and Muse Wonder Full, with one headline claim: active memory that persists across days, weeks, and months. Pricing is $4 per million tokens in, $20 per million out, which puts it squarely at the same rate as OpenAI and Anthropic. At identical pricing with no switching costs, memory is the only reason to move.

Whether that reason holds depends on what "active memory" actually is. Genuine continual learning, where the model updates its weights from new information without forgetting old information, has been an open research problem for a decade. If that is what Meta shipped, it is the first genuinely new capability from the company in two years. If it is retrieval scaffolding, meaning past conversations stored in a database and fed back in at inference time, then every competitor already has the ingredients and this is a branding exercise. Meta has not published a paper or provided independent provenance for MemoryBench, its own benchmark. That is not reassuring. A fast test: feed Wonder contradictory facts about the same user across separate sessions and check whether it reconciles them or regurgitates the most recent. That tells you scaffolding from real memory in an afternoon.

The paywall on Wonder Full tells a separate story. The better model is locked to Muse Code subscribers and off the API entirely. The only model a developer can actually build on is Wonder, the second-tier option. Meta would not structure it that way if the economics on Wonder Full worked at open API rates. Serving a frontier-tier model at $20 per million output tokens is expensive, and a captive subscription provides predictable revenue that an open API does not. That is a confession about serving cost, not a product decision.

For enterprise buyers, Wonder Full is dead before procurement starts. No SSO, no audit logs, no data-residency guarantees, no indemnification, and memory raises its own legal questions. A model synthesizing a user's emotional patterns over six months is a materially different product from a stateless assistant, and Meta carries active FTC consent-decree obligations on user data handling. The announcement says nothing about memory audit trails, deletion rights, or adversarial memory poisoning. Healthcare and finance teams will block it on day one.

The call: Muse Wonder Full will still not be available through the Meta Model API by July 13, 2027. Inference cost does not fall fast enough in nine months to flip the math, and Meta has no competitive pressure forcing it onto unprofitable pricing. If Wonder Full reaches the API before then, it will require either a steep serving-cost drop or a price increase, and nothing in the release points to either.

The subscriber lock is the whole story for builders. You cannot build on the model Meta is actually proud of.

Meta shipped two frontier-tier models, Muse Wonder and Muse Wonder Full, and the pitch is "active memory": a model that remembers you across days, weeks, months. Pricing is $4 per million tokens in, $20 per million out, which is ordinary. The thing worth deciding on is whether the memory claim is real and whether you'd build on it.

How hard is this to undo? Easy. Nobody has to migrate. Wonder sits on the Meta Model API next to OpenAI and Anthropic, and you can test it this week and walk away Friday with nothing lost but some eval time.

What's actually being decided: not "do I switch to Meta" but "is persistent memory a feature I now have to build around, and does Meta's version work well enough to trust with real user data."

What sets the deadline: nothing. No shutdown, no price cliff, no contract. The only clock is competitive, meaning whether OpenAI and Anthropic ship their own memory story and make this normal.

---

The Skeptic. Meta has shipped "competitive" frontier models before and the gap with OpenAI widened anyway. MemoryBench is Meta's own benchmark, which makes a win on it worth roughly nothing until somebody outside the building reproduces it. The restructure under Alexandr Wang is an org story. Scale AI's trade is data labeling, not model architecture, and a reshuffled lab shipping a frontier model is a tidy redemption arc that flatters the evidence. The paywall on Wonder Full tells you plenty: Meta won't expose its best model to developers who would immediately run cost-per-task comparisons against GPT-6 and Fable 5.1. If the economics were clean, the full model would be on the API.

The Safety Lens. A model that synthesizes your emotional patterns over six months is a different animal from a stateless assistant. That is exactly the capability regulators and alignment people have been flagging, and Meta is the worst-positioned company to ship it quietly. The FTC consent decree over user-data handling is still active. The announcement says nothing about memory audit trails, user deletion rights, or what happens when someone poisons the memory with bad inputs on purpose. "Active memory" is the headline feature and the biggest legal exposure in the same breath. Anyone in healthcare or finance will block it on day one until there's documentation on how memory is stored, versioned, and wiped on request.

The Researcher. The only technically new claim here is active memory, and it deserves rough handling. Continual learning without catastrophic forgetting, where a model learns new things without wiping out what it already knew, has been an open problem for a decade. Calling it solved via a proprietary eval is a red flag. "Days, weeks, months" could mean genuine in-weights memory or it could mean retrieval scaffolding: storing your past chats in a database and feeding them back in. Those are wildly different things with wildly different reliability. No paper, no independent MemoryBench provenance, no contamination controls. Publication or it didn't happen.

The Enterprise Buyer. Wonder Full is locked to Muse Code subscribers and off the API entirely. That kills it for most procurement before the conversation starts. A CTO cannot sign for a model they can only reach through a consumer subscription, with no SSO, no audit logs, no data-residency guarantee, and no indemnification. So the buyable product is Wonder, the second-best model, at pricing identical to OpenAI and Anthropic. There is no reason to switch an existing integration. Memory would be the reason, but memory is the part legal will not clear without deletion-rights documentation that does not yet exist.

The Compute Pragmatist. Prometheus in Ohio is the first internal cluster Meta has publicly bolted to a frontier release, and that is a real capex signal. The Wonder / Wonder Full split is inference economics talking. Serving the full model at $20 per million output tokens on the open API would bleed money, so Meta parked it behind a subscription where revenue is captive and predictable. That's a confession about serving cost, not a product-tiering strategy. If Wonder Full never reaches the API inside 18 months, the full model was always too expensive to serve at market rates.

---

The personas collide in two places. The Researcher wants a MemoryBench result nobody can reproduce, while the Builder could ship session-persistent agents on it tomorrow if the memory held. One cares whether it's real science; the other cares whether it works at 10,000 requests. Both are stuck on the same missing thing: documentation on how memory actually functions.

The Compute Pragmatist and the Skeptic read the paywall the same way from opposite ends. Skeptic says Meta is hiding the best model from cost comparisons. Pragmatist says Meta can't afford to serve it on the open market. Same fact, same conclusion: Wonder Full's economics don't work at API prices, and the subscription is the subsidy.

This hinges on one belief. Is active memory genuine in-weights continual learning, or is it retrieval scaffolding with a marketing name? Everything else, the pricing, the org story, the data center, is noise. If it's scaffolding, every competitor already has the ingredients and this is a branding exercise. If it's real continual learning without forgetting, it's the first genuinely new capability from Meta in two years. Before building anything on it, run one test: feed Wonder contradictory facts about a user across sessions and check whether it reconciles them or just regurgitates the most recent. That tells you scaffolding from memory in an afternoon.

Prediction: Muse Wonder Full will still not be available through the Meta Model API by 2027-07-13, nine months after launch.

Confidence: Medium. The paywall reflects serving cost, and serving cost does not fall fast enough in nine months to change the math.

Why: Meta locked its best model to Muse Code subscribers while putting the weaker Wonder on the open API at the same $4/$20 pricing as OpenAI and Anthropic. The split only makes sense if Wonder Full costs too much to serve profitably at those rates, so Meta needs the captive subscription revenue to subsidize it. Inference cost for a frontier-tier model does not drop enough in nine months to flip that math, and Meta has no competitive reason to rush an unprofitable model onto the API when rivals aren't forcing its hand. The opposite outcome, Wonder Full hitting the API soon, would require either a steep serving-cost drop or a pricing bump, and nothing in the release points to either.

Revisit by 2027-07-13: We're right if Muse Wonder Full is still subscriber-only and absent from the Meta Model API on that date. We're wrong if Meta lists Wonder Full as a programmatically callable model on the Meta Model API before then.

The subscriber lock is the whole story for builders. You cannot build on the model Meta is actually proud of. You can only build on the one it was comfortable pricing against the field.