Meta's Muse Wonder AI Model Puts Its Best Safety Features Behind a Paywall
Meta's newest AI agent model is free to try, but the strongest protections against misuse only come with a subscription.
Meta put its newest AI model out on Monday, and it's called Muse Wonder, codenamed Watermelon internally. Alexandr Wang, who runs Meta's AI shop, says it's built to power "personal and professional agents," the kind of AI you'd ask to book a dinner reservation or check on availability without you having to sit there and walk it through every step. The pitch is that when the agent hits a roadblock, it tries another approach instead of stalling out and running back to you for help. That sounds small, but it's the difference between an assistant you can actually trust with a task and one you have to babysit.
Here's the catch. Muse shot to the top of Apple's App Store after Meta launched the Muse agent app last month, and it's free to download. But the real safeguards, the ones Meta says protect against "cyber misuse," only come with the paid version.
Why the Safety Features Come With a Price Tag
Wang says Meta's team has been watching for "incidents or issues" with Muse since it launched. Some users on social media said the agent got into their personal messages without being told to. Meta disputes that it happened without authorization, and Wang's explanation is that people may have given the agent permission to do something earlier in a conversation and then forgotten about it days later. Meta says it's working on educating users so they understand what they've already agreed to.
That is worth sitting on for a second. If the strongest protections against misuse are locked behind a subscription, the free version of an agent that can read your messages and book things on your behalf is, by Meta's own description, the less protected version.
What an Agent That Does Errands Means for the Jobs It Replaces
An AI that can handle reservations, scheduling, and "various setbacks" on its own is aimed squarely at the kind of coordination work that used to need a human assistant, a receptionist, or a scheduler. That is not a huge chunk of the job market by itself, but it adds to the pile of entry level coordination and admin jobs getting absorbed by AI tools, on top of what we have already seen this year. If you are in a job where your value is mostly keeping other people's calendars straight or handling routine back and forth, this is one more signal to build a skill on top of that work that an agent cannot do yet, like judgment calls, relationships, or actual expertise.
The Liability Question Small Businesses and Insurers Will Have to Answer
This is where my insurance background makes me look at this differently than most people reading the headline. If an AI agent is booking things, accessing messages, and making decisions for you or for your business, somebody has to own the risk when it gets something wrong, books the wrong thing, or exposes something it should not have. Right now that liability question is unsettled. For a small business owner handing scheduling or customer communication over to an agent like this, I would want clear answers on who is responsible if the tool makes a mistake that costs a customer money, before I built it into how I run things. In insurance, this is exactly the kind of gap that shows up at renewal time when a carrier asks what tools you are using and whether you have controls around them.
Why Coders Will Pay More for This One
Muse Wonder is also going to coders through Meta's Model API, and Wang said it will cost more than the standard Muse subscriptions because it is a bigger, frontier level model. He also said it will still be priced below what OpenAI and Anthropic charge for their top models. For anyone running a small software or marketing shop, that pricing gap matters. If you build tools or client work on top of AI models, the cost of the model you pick is a real line item, and Meta is betting it can win business by being the cheaper frontier option rather than the free one. That is worth watching if you are deciding which model to build your business on.
What to Watch Before You Let an Agent Run Your Business
If you use Muse or a similar agent for your business, read exactly what permissions you are granting before you grant them, not after something goes wrong. Check whether your current subscription tier actually includes the "cyber misuse" protections Meta is advertising, since those are apparently not standard. And if you are a coder or small shop evaluating the Model API, compare the actual price per task against OpenAI and Anthropic before you commit, since Meta's advantage is supposed to be cost, not just capability.