Muse Wonder Explained: Meta's New AI Models Behind Your Agent

When Meta launched Muse on September 8, 2026, it described the product as a personal AI agent rather than a chatbot — software that completes real multi-step tasks inside a secure virtual computer. A little over a month later, Meta showed the engine that will drive that agent going forward. On October 12, 2026, the company unveiled the Muse Wonder series: two frontier AI models designed for exactly the kind of long, complicated, personal work an agent like Muse is supposed to do.

What Muse Wonder is (and what it isn't)

Muse Wonder is the successor to Meta's Muse Spark line. Muse Spark, introduced in April 2026, was built for fast, lightweight responses, with Spark 1.3 as its most capable version. The Wonder models are larger and far more ambitious: Meta calls them its first true frontier releases, aimed at personal and professional agentic tasks — from insurance claims to software builds.

The headline spec is a 1 million token context window, holding entire projects in view at once, plus text, image, video, and audio input. Meta says the models were trained at its gigawatt-scale Prometheus data center on long, realistic runs — multi-file coding projects, repeated testing and revision — so they can sustain progress on work that takes real agents hours or days.

On benchmarks, keep your skepticism on. Meta claims state-of-the-art results, at times beating Claude Fable 5.1 and GPT-6 Astra. Independent analysis from Artificial Analysis scored Wonder Full at 56 and Wonder at 52 on its Intelligence Index, ahead of Muse Spark 1.3's 48. But most headline coding numbers come from Meta's own runs under setups that may not match rivals' — treat them as directional until independent testing lands.

Active Memory and Active Focus, explained

The names are marketing, but the ideas behind them are straightforward. Active Memory is the model's ability to carry important details about you across conversations and pull them back up when useful. Meta designed it to remember preferences and key facts over days, weeks, even months — the kind of context an assistant would pick up naturally. Meta measures it with a benchmark called MemoryBench and claims Wonder beats rival models; that claim is Meta's own, and independent verification is pending.

Active Focus is the other half: sticking with a hard goal instead of stalling. It tries alternative approaches, works through blockers, adapts as new information arrives, asks for clarification when genuinely ambiguous, and checks with you before important actions. Meta evaluates this with FocusBench and a long-running software test called SWE-Marathon.

Keep them apart this way: Active Memory is about knowing your context; Active Focus is about finishing hard jobs. Ask Muse to plan a two-week trip, and Memory means it remembers your dates and budget without you repeating yourself; Focus means that when the third flight search fails, it tries a different route and asks before touching your credit card. One is continuity; the other is stamina.

Who gets Muse Wonder, and how

Right now the models are aimed at builders, not app users. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full — the larger variant with the strongest scores — is limited to Muse Code subscribers.

Most Muse users care about what's next. Meta said Muse Wonder will "soon power experiences in our personal agent product, Muse, rolling out first to subscribers." Zuckerberg was blunter: subscribe today to get the Wonder-powered version. That means the $20 and $100 tiers see Wonder's smarts first — for artifact generation and the hardest tasks — while free-tier users wait.

On pricing: $4 per million input tokens and $20 per million output tokens, versus $1.25/$4.25 for Muse Spark — a sharp jump to frontier-model pricing. And Wonder is hosted-only, with no downloadable weights: a deliberate break from the open-weights Llama era, signaling Meta's direction toward a paid, tiered agent platform.

What this means for everyday Muse users

If you already use Muse, Wonder is the upgrade that makes the agent feel more like a real assistant. The practical wins: less repeating yourself across sessions, errands that run end-to-end without babysitting, and an agent that pauses to confirm before spending money or sending messages — behavior Muse already promises, now backed by a model trained for it.

Meta calls Wonder its most aligned model to date, trained to protect private information and resist malicious instructions — essential for an agent that can read your messages, calendar, and documents. Critical press has described earlier test failures and claims Meta disputes; review what your agent can do before letting it run unattended.

One reality check: Muse remains US- and Canada-only (Canada was added on September 18, 2026), with no third country announced. The Wonder rollout is a model upgrade for eligible users, not an expansion of who can sign up.

Is Muse Wonder available in the Muse app right now?
Not yet, as of mid-October 2026. It was announced on October 12 for developers via the Meta Model API and Muse Code, and Meta says it will "soon" power the Muse personal-agent app, rolling out first to subscribers of the $20 and $100 tiers.

What's the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder Full is the larger, more capable model in the series — 56 on the Artificial Analysis Intelligence Index versus 52 for Muse Wonder. Full is currently reserved for Muse Code subscribers.

How much does the Muse Wonder API cost?
$4 per million input tokens and $20 per million output tokens — roughly three times the pricing of the previous Muse Spark generation ($1.25/$4.25). There is also a $0.20 per million cached-token rate reported by third parties, though that figure is not officially confirmed.

Can I download Muse Wonder's weights like Llama?
No. Muse Wonder is hosted-only — available through the Meta Model API and Muse Code — with no downloadable weights. That is a deliberate break from the open-weights Llama approach.

If you are watching Muse from outside the US and Canada, or just want to be ready when the Wonder-powered rollout reaches your tier, get your setup prepared now: a Meta account, being 18+, and a clear understanding of how the product works. Our Muse account registration assistance walks you through registering your own Meta account for Muse, step by step — help only, per platform terms: we help you set up your own account and never deal in buying or selling accounts. And our guide on preparing your own Muse account covers the pre-sign-up checklist, so you can try the Wonder-era Muse the moment Meta opens the door.