Meta has shipped its first model built to compete at the top of the field. Muse Wonder and Muse Wonder Full arrive with agent-friendly features, a price tag close to the current leaders, and benchmarks the company hasn't yet opened to outside testing.
What Meta Actually Released With Muse Wonder
Meta's Superintelligence Lab put out two models on Monday: Muse Wonder and Muse Wonder Full. The company calls them its most capable yet and its first real swing at what the industry calls the frontier, the top tier of AI models that currently includes OpenAI's GPT-6 line and Anthropic's Opus family.
The two differ in who can use them. Muse Wonder is available today through the Meta Model API and through Muse Code, the company's coding tool. Muse Wonder Full is the stronger sibling, and it's locked to Muse Code subscribers. It won't be served through the API at all, at least for now. Both models are also meant to power Muse, Meta's personal assistant app.
That split matters if you build on top of these models. Paying API customers get the smaller of the two, while the most capable version stays inside Meta's own products.
The Active Memory Pitch Behind Muse Wonder
Meta spent most of its launch post on one capability it calls Active Memory. The pitch is simple: a model should remember what matters to you across chats, days, and weeks, then pull the right details back when a task needs them.
It's a real gap in everyday AI use. Anyone who has pasted the same context into a chatbot for the tenth time knows the friction. Meta's research scientist Kavya Kopparapu framed it as the point of the release: the effort went into making the model remember what's important "over days, weeks, months, and beyond."
There's a second piece called Active Focus, which is about finishing hard tasks. Meta says the model keeps working when a job gets difficult, tries other approaches, recovers from errors, and asks for clarification instead of stalling. For an agent that's supposed to handle an insurance claim or read through a lease, that persistence is the difference between a demo and something you'd actually rely on.
Muse Wonder's Benchmarks Are Meta's Own Numbers
Meta says Wonder and Wonder Full score in the same range as Fable 5.1 and GPT-6 Astra on its internal tests. Those results come from Meta's own model card. No outside lab has published an independent run yet, so treat the comparison as a claim, not a settled ranking.
Continuing learning and memory frameworks were tested with a benchmark called MemoryBench, and Meta reports Muse Wonder landing above Fable, Opus, and Astra there. The same caveat applies. Vendor-run benchmarks show what a company chose to measure, and independent evaluators like Artificial Analysis were still working through their own numbers when the models went live.
Here's the practical read: if remembering your context over time is the feature you care about, the only test that counts is trying it on your own work.
What Muse Wonder Costs Through the API
For developers using the API, Muse Wonder runs $4 per million input tokens and $20 per million output tokens. That puts it roughly level with GPT-6 Sol and Opus 5.5 on price. You're not getting a discount for switching to Meta.
Model
- Muse Wonder
- Muse Wonder Full
Access and price
- Meta Model API and Muse Code, $4 in / $20 out per million tokens
- Muse Code subscribers only, no public API
The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year. The launch follows a 14-month rebuild of Meta's AI efforts that started after Scale AI founder Alexandr Wang restructured the division. Before this, Meta had shipped the lightweight Muse Spark and the Muse agent but had stayed out of the top-tier model race that OpenAI and Anthropic dominate.
Is Muse Wonder Worth Switching Your Assistant For
If you already pay for a frontier assistant, Muse Wonder's first job is to prove it remembers you better than what you use now. The API price gives you no reason to move on cost alone, and the strongest version isn't available to API users at all.
For anyone building agents around memory-heavy tasks, Active Memory is a specific claim worth testing directly. Run a week-long project through both and see which one still holds the thread.
Muse Wonder is live now through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers, and the company says more of Muse will run on the new family soon. Watch for the independent benchmark results over the next few weeks, since that's when Meta's own numbers get a reality check.
