For years, Meta was the big tech company that gave its AI away. Its Llama models went out the door for anyone to download, tweak and run on their own servers. So here’s the twist. Meta’s most powerful model yet, Meta Muse Wonder, is staying locked up.
Meta released Muse Wonder and a bigger sibling, Muse Wonder Full, on 12 October. TechCrunch called it the company’s first frontier-class release, meaning a model built to sit at the very top of the field. And it tells you a lot about where Meta’s head is at right now.
So what actually launched?
Both models come from Meta Superintelligence Labs, the unit behind Meta’s AI reboot. In Meta’s own announcement, the company says, “Fourteen months ago, we began an overhaul of our AI efforts.” The new pair builds on Muse Spark, a lighter model Meta launched in April.
Meta pitches two headline features. The first is Active Memory, which it describes as “the ability to carry what matters to you across conversations.” The second is Active Focus, “the ability to keep working toward your goal when a task gets hard.” In practice, that means trying another approach when something fails, then checking in before important actions.
The examples are very everyday. Meta talks about sorting out an insurance claim, understanding a lease or making sense of your finances. It also claims frontier-level coding, like building an app for a new business. That’s classic AI agent territory.
Developers pay $4 per million input tokens and $20 per million output tokens on Meta’s API. A token is a chunk of text, roughly a short word. Wonder Full isn’t on the API at all, though. You only get it through Muse Code subscriptions, at $50 or $200 a month.
Wait, Meta went closed?
Yep. Neither Wonder model comes with open weights, as Futurum analyst Nick Patience points out in his breakdown. Closed weights means you can’t download the model itself. Instead, you rent access through Meta’s servers.
That’s a real break from the Llama era. Companies liked Llama because they could host it themselves, fine-tune it on their own data and keep their prompts off someone else’s cloud. Wonder, as a hosted service, offers none of that.
To be fair, Meta hasn’t abandoned open models entirely. It still offers Muse Glimmer, an open 30-billion-parameter model that came out in August. But the best stuff now sits behind a paywall.
Honestly? We get it. Meta trained these models at Prometheus, its gigawatt-scale data center in Ohio. Giving away something that expensive is a tough sell to shareholders.
How good is Meta Muse Wonder, really?
Pretty good, but not the champ. On the Artificial Analysis Intelligence Index, which rolls ten evaluations into one score, Wonder Full scored 55.8. That puts it ahead of OpenAI’s GPT-6 Astra (52.7) and Google’s Gemini 4 Argon (52.6). Yet it still trails Anthropic’s Claude Opus 5.5 (57.6) and Sonnet 5.5 (56.0). The standard Muse Wonder landed at 52.4.
Here’s the thing, though. The top six models sit within about five points of each other. So “best model” is starting to matter less than price, speed and fit for your actual work.
Meta’s favorite number is about memory. On MemoryBench, Meta’s own test, Wonder scored 85.6% while Opus 5.5 and GPT-6 Astra came in around 81%. Since Meta picked the test, take that one with a pinch of salt.
The part that made us pause
Memory is great until it isn’t. An assistant that remembers your lease, your money and your calendar is useful. It’s also a lot of personal data sitting in one place.
Meta’s own safety chart shows Wonder mishandled personal data in 3.4% of test cases. So that’s roughly one failure in 30. Patience called that rate “a hard sell to enterprise risk committees,” and we’d add regular users to that list.
Meta also says it trained the models to avoid sharing private information without permission. Its model card admits it can’t rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse, though. A fuller safety report is on the way, according to Meta.
Follow the money
Why build this at all? Look at Meta’s revenue. Fortune notes that Meta brings in about $201 billion a year, and upwards of 98% comes from digital ads. That’s a lot of eggs in one basket.
Agents could become a second basket. Deutsche Bank analyst Benjamin Black suggested Meta’s personal AI agent could reach up to 8% of revenue by 2030, roughly $36 billion. Meta also has reach on its side, because the Muse app passed 3.4 million downloads through 24 September, based on Sensor Tower estimates.
Okay, but should you care?
If you build software, yes. Meta Muse Wonder is one more serious option at a mid-range price, with a 1-million-token context window. Test it on your own tasks before you switch, because a cheap token price doesn’t always mean a cheap finished job. Our guide to where LLMs do real work is a good place to start picking those tasks.
For the rest of us, including the AI crowd here in Armenia, the bigger story is the shift. One of the loudest champions of open AI just kept its best model closed. That matters if your startup’s plan depends on free, top-tier models you can run yourself.
Our take? Meta Muse Wonder proves Meta can build at the frontier. Whether people will pay for it, and trust it with their memories, is still the open question.