Meta did not win the October model race on price or on peak score. On the independent Artificial Analysis Intelligence Index that VentureBeat published alongside the launch, Muse Wonder Full lands at 56, under Claude Opus 5.5’s 58, roughly level with GPT-6 Astra at 53 and Gemini 4 Argon at 53. The cheaper Muse Wonder scores 52. Pricing matches, rather than undercuts, the field: $4 per million input tokens and $20 per million output on Wonder, the same rate as Opus 5.5, well under Astra’s $10/$50.
So the thing Meta chose to put at the center of its first frontier model, announced October 12, is not a benchmark win. It is memory. And memory is the part of this launch that changes what you are actually buying.
The pitch is a model that keeps your context
Meta’s own announcement builds the Muse Wonder series around two features. Active Memory is described as “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.” Active Focus “continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information,” and checks before consequential actions. A Meta research scientist put the goal plainly to TechCrunch: “We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond.”
That is a real differentiator. Every frontier assistant today starts most sessions cold, or leans on a thin retrieval layer bolted on top. A model trained to carry your preferences, your project state, and your prior instructions forward, without you re-explaining them, is a genuinely different product shape. Meta backs it with a one-million-token context window, text, image, audio and video input, and rate limits of 3,000 requests and 4 million tokens per minute. The capability is the easy part to admire.
A model built to remember you is a data store, not a feature
Here is the reframe the launch coverage skipped. A model whose selling point is that it remembers you is, by construction, a standing record of your personal and business context. It is not a toggle you flip on. It is a place your data now lives, retained across sessions, retrieved into future work, and held somewhere you did not choose.
Meta knows this, and its engineering note says so in the one passage worth reading twice. Effective memory, the company acknowledges, has to solve governance problems that stateless models never had: “A remembered preference from an earlier conversation should not automatically override a new instruction. Similarly, information associated with one customer or project must not be exposed in another user’s session.” That second sentence is the whole risk category. Cross-session memory is also cross-session leakage waiting to be mis-scoped, and Meta is naming it on launch day because it is the first thing a serious buyer asks.
I spent twenty years in IT operations, and the question that memory feature raises is one records management has answered for decades: what does this system retain, where does it keep it, under what retention schedule, and how do I make it forget. A model that remembers your deal terms and your customer notes belongs on the same data-classification and retention register as any application that holds that information, the same way a connected AI agent is a machine identity you eventually have to offboard. “Can I delete what it learned about this client” is not a feature request. It is a retention question, and it should have an answer before the first prompt, not after the first incident.
The best model is coupled to a data-harvesting default
The governance problem gets sharper because of how Meta gated access. The headline model, Muse Wonder Full, has no public API. You can only reach it through a Muse Code subscription, which runs $50 to $200 a month. Plain Muse Wonder is the one available for pay-as-you-go API use.
Now layer the data modes on top. Meta offers three. Standard mode does not use your submitted content for training, and it is the default on the Wonder API. Contributor mode “permits Meta to use submitted prompts, code, files, contextual information and outputs to improve future models,” and it is the default for Muse Code subscriptions. Zero Data Retention exists for qualified organizations, but turning it on restricts conversation state, file uploads, and web search.
Read those three facts together and the coupling is hard to miss. The best version of the model is reachable only through the subscription whose default lets Meta train on your prompts, your code, and your files. And the one setting that would stop ordinary retention, Zero Data Retention, also disables the conversation state that the memory feature depends on. In other words, the privacy control and the product you came for are pointed in opposite directions. You can keep the memory, or you can keep the data boundary, and the default ships with the boundary down. That is a different bargain than Anthropic’s move to make 30-day logging the floor with ZDR available on top, and worth pricing as such.
None of this is hidden; it is in the service terms. But a default that harvests is still a default, and defaults are what teams actually run. The Contributor setting is a checkbox someone has to find and change on every Muse Code seat, on the one model whose entire purpose is to accumulate a long memory of your work.
The scores you cannot check, and the ones that missed
Meta’s reported head-to-heads look strong in isolation: Wonder Full at 79.0 on DeepSWE 1.1 against Astra’s 74.1, 57.5 on the long-running SWE-Marathon against Opus 5.5’s 46.3, and a wide 13.1 to 4.2 lead over Opus on the Harvey legal-agent benchmark, all per VentureBeat’s rundown. Those are Meta’s own numbers, and the picture outside its slide deck is mixed rather than dominant. Wonder Full wins some comparisons and loses others, trailing Opus 5.5 on the overall Intelligence Index and on several coding tasks.
Two problems make the scores worth discounting. First, early hands-on testing surfaced the usual benchmark-to-reality gap: users reported Muse Wonder generating red cylinders when asked for a lava lamp effect, the kind of ordinary task where a frontier score should not wobble. Second, and more structurally, the model carrying Meta’s headline numbers is the one you cannot independently test, because Wonder Full has no public API. A top-line result you cannot reproduce on your own workload is a marketing claim, not a measurement, which is exactly why the eval that predicts what ships has to be built on your own tasks, not read off a vendor chart.
Where it fits, and the three questions to ask first
Muse Wonder earns an evaluation slot if you are doing memory-heavy assistant work, personal or professional, where carrying context across days actually saves rework, and your data is a mix of public and low-sensitivity internal material you are comfortable running on Meta’s stack. The price is competitive, the context window is large, and the memory behavior is the first mainstream attempt to make persistence a first-class capability rather than an add-on.
It does not earn a production commitment today for regulated, data-residency, or strict-provenance work. Neither Wonder nor Wonder Full ships with downloadable weights, so the self-hosted path that answers sovereignty requirements is closed here, and the memory feature pulls against Zero Data Retention. As with Google’s membership-gated Gemini 4 Argon, you are choosing an access and governance regime, not just a model, and Meta’s regime defaults toward keeping your data.
Before you connect anything beyond a test account, get three answers in writing. What does Active Memory retain, where is it stored, and under which service mode is each seat running. Can you force-delete what a model has remembered about a specific client or project, and how fast. And is Contributor mode off on every Muse Code subscription, confirmed in the console, not assumed from the brochure. The capability is the headline. What it keeps is the decision, and that one belongs to whoever owns your data on the risk register, not to whoever liked the demo.