This week, we had an initial rollout of ground-breaking new models from the company formerly known as Facebook, and its head, Mark Zuckerberg, who has famously invested billions in AI.

Now, along with the Meta Ray-Bans, which I haven’t heard much about recently, Meta has these consumer-facing public models with subscription plans attached. So what’s under the hood?

Active Memory and Active Focus

These features, touted in the literature around Muse Wonder, are meant to facilitate “sustained work.” A 1-million-token context window helps.

“The clearest fit is anyone whose immediate need is agentic reasoning or coding: builders who want a capable model for long-horizon, multi-step tasks that hold context across a thread,” writes a reviewer at Kompozy . “(Active memory) retains relevant context across conversations and recalls it when needed — a genuine step up for long, continuous work.”

That overview of the model’s depth makes sense, given the need for systems that are stateful and agentic.

Muse Wonder also excels at sitting down to a laptop screen in a way that parallels human user activity: the models can operate browsers, desktop, and mobile apps. Muse Wonder, Meta says, “plans long tasks, delegates to sub-agents, and recovers from errors.”

That apes the capabilities of Anthropic Claude models, and also hews toward that agentic situation that everyone is talking about. This means you can get Muse the resources, turn it loose, and let it “tap” away, figuratively, that is.

“For years, Meta has seemed inexcusably late to the AI party,” writes Thomas Smith at Fast Company. “It was tempting to write Meta off as an also-ran. Perhaps Mark Zuckerberg was still gun-shy from the abject failure of the Metaverse, and was going to sit this particular revolution out. Then, Muse landed. Almost overnight, the platform was stealing market share from Meta’s competitors, and growing shockingly fast. Less than a month post-launch, Muse had racked up 6.6 million downloads, and an even-more-impressive 1.8 million daily active users. Citi thinks it could generate $27 billion in revenue in fairly short order. Meta, in other words, was late to the party—but now that it’s arrived, the company’s tech is all anyone can talk about.”

It’s also telling that Muse Wonder is getting high marks on a range of tests meant to evaluate models.

Reviews note the following scores on Artificial Analysis’s Intelligence Index: Muse Wonder Full gets a 56 — tying Claude Sonnet 5.5 and beating GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1, each of which scored a bit lower. The standard Muse Wonder reportedly receives a 52.

Also, check out SWE-bench stats and more.

Here’s the big caveat on Muse Wonder:

The company notes that this model is omnivorous, to an extent, on training and input data, but only spits out results in written form. No video. No image. Muse Wonder returns its findings in the form of “plans and documents.”

Therefore, this model is great for some use cases, and terrible for others. If you’re a paralegal, fantastic – the last thing you want is to get a cartoon or a funny video. If you’re looking to create an infographic, or wow the crowd with an explainer reel, you’re at the wrong place. To be fair, since OpenAI pulled Sora, there’s not really a front-running video gen model out there – there’s Runway, Synthesia, Pictory AI, Klap, Descript, but hardly anyone knows what those are.

As for pricing, insiders note that at $4 per million tokens in, and $20 per million tokens out, Muse Wonder has the same pricing structure as Claude, currently, where Muse Wonder Full is only available with an upper tier subscription.

At the end of the day, the new Meta models seem to have some amount of staying potential. Go out and check these out for yourself, as 2026 winds down.