Is Muse Wonder Safe? Meta’s New Model, Memory and Your Data
Meta released Muse Wonder on 12 October 2026. What Meta documents about its data use and Active Memory, what it has not published yet, and how to start.
Short answer: nobody outside Meta can say yet. Meta released Muse Wonder and Muse Wonder Full on 12 October 2026, and its launch post says a full report with additional safety evaluations will be "available soon". Until it lands, what you can judge is narrower: where your data goes on each way of using the model, and what the new memory feature means for what you type into it.
This post sticks to what Meta has written down, as of 13 October 2026. It separates the things Meta documents from the things it has not said, and it ends with a short way to try the model without handing it more than it needs.
What Meta released, and where it runs
Muse Wonder is a model, not an app. That matters, because Meta now has several things called Muse. Muse is the personal agent app that launched in the US in September. Muse Spark is the model family that has powered Meta AI since April. Muse Code is Meta’s coding agent. Muse Wonder sits above Spark and, by Meta’s description, is its most capable model to date. The difference between the app and the models is covered in what Muse Spark is.
TechCrunch calls it Meta’s first attempt at a frontier model. There are two versions, and they reach you in different ways:
- Muse Wonder is on the Meta Model API as muse-wonder-1.0, with a 1 million token context window. Meta’s price list shows $4 per million input tokens, $0.20 cached and $20 per million output tokens, in US dollars. It is also in Muse Code.
- Muse Wonder Full is not on the API. Meta lists it as available by subscription in Muse Code only.
- Both will later power the Muse app, rolling out first to subscribers. Meta gives no date for that.
Meta’s announcement names no countries. The Model API page itself only says "expanded global access". If you are outside the US, check availability in your own account rather than assuming it from a launch post.
Does Meta train on what you send it?
It depends on the route, and this is the part worth getting right before anything else. Meta’s Model API price list labels each model variant with how your data is used. The cheap "contributor" variants of Muse Spark are labelled "Used to improve our products". The standard variants are labelled "Not used to improve our products".
For Muse Wonder on the API, no contributor variant is listed. There is one price, the standard one. What the price list does not do is print a data-use label next to muse-wonder-1.0 the way it does for Spark. The reasonable reading is that Wonder sits on standard terms, but that is a reading of a price table, not a sentence from Meta. If it matters for your work, ask Meta in writing.
Muse Code is the route to watch. Its plans page runs from $5 to $200 a month and shows both kinds of variant, contributor and standard, with the same two labels. Wonder Full is only reachable through those plans. So before you paste client code or a confidential document into Muse Code, find out which mode your plan is in and whether you can switch it. The trade between cheap and private on Muse Code is set out in Muse Code vs Claude Code.
Active Memory: useful, and a reason to be careful
The new feature in Muse Wonder is what Meta calls Active Memory. Meta’s research post describes it as carrying important context about you across conversations and retrieving the relevant details for the task in front of it.
That is genuinely handy. You stop re-explaining your job, your clients and your preferences every time. It also changes the habit you need. With a model that forgets, a careless paste lasts one conversation. With a model built to remember, it can come back weeks later in a different task. Meta’s pages do not say where those memories are stored, how long they are kept, or how you review and delete them. Those are the three questions to answer before you rely on it.
On private information more broadly, Meta says it trained the Muse Wonder family "to avoid sharing private information without permission". That is a statement about the model’s behaviour, not a promise about storage. Both matter, and only one of them is on the page.
What Meta says about safety, and what it has not shown
Meta calls Muse Wonder its most aligned model to date. It says the models are trained to decline requests to help with harmful acts, to resist adversarial instructions, and to carry Meta’s strongest safeguards so far against chemical, biological, radiological and nuclear misuse and cyber misuse. It also says it kept the models useful for defenders, so security teams can still use them to find flaws in their own code. All of this is under what Meta calls its Superintelligence Scaling Framework.
What is missing is the evidence. The detailed evaluations are in the report Meta says is coming. Compare that with how other launches have been handled: some vendors publish a long system card on day one, and their safety posts can quote it, as is GPT-6 Astra safe does. Here the honest position is to wait for the report before trusting Wonder with anything where a mistake is expensive.
The other gap is the ordinary one. Meta pitches Wonder for law, finance and health work. A model that is strong on benchmarks still gets things wrong with confidence, which is the pattern in what AI is actually bad at. Expert-sounding output in a field you do not know is exactly where that bites.
A careful way to start
- Pick the route on purpose. The API and Muse Code have different data terms. Decide which one you are on before you decide what to paste.
- Check your Muse Code mode. If your plan uses a contributor variant, treat everything you type as something Meta may use to improve its products.
- Keep work data out until your employer says yes. Client files, contracts and colleagues’ details are not yours to send to a new service.
- Watch what it remembers. Tell it less than you would tell a colleague until Meta documents how to review and delete memories.
- Check the answers that matter. Use the routine in how to check an AI answer when you are not the expert before anything goes out under your name.
- Read the safety report when it lands. It is the document that turns "Meta says" into something you can weigh.
If your organisation wants a structure for deciding, the NIST AI Risk Management Framework boils down to three plain questions: what can this system touch, what is the worst realistic mistake, and what have we done to limit it?
The short version
Choose the route before the prompt, keep work data out, and wait for Meta’s safety report before trusting Wonder with anything expensive.
If you are weighing the Muse app rather than the model, the questions are about connected accounts and approvals instead, and is Meta Muse safe covers them. For the cost side of Meta’s line-up, see Meta Muse pricing.
New models arrive every few weeks now, and the useful skill is the same each time: know where your data goes, check what comes back, and keep the risky jobs in human hands. Coursium teaches people to use AI at work, on their phone, with short lessons on exactly that. You can see how Coursium works if that is useful.