Meta has introduced two new AI models, Muse Wonder and Muse Wonder Full, describing them as a major step toward what it calls personal superintelligence. The headline capabilities are not just about generating an answer to a single prompt. Meta says the models add active memory and active focus, two features meant to make an AI agent more useful across extended, multi-step work.
That distinction matters. Plenty of AI interactions begin and end with one request: summarize a document, draft an email, explain a piece of code. Meta’s framing for Wonder is instead centered on continuity. Its planned destination is Muse, the company’s personal AI agent, beginning with paying subscribers. The promise is an assistant that can retain relevant details across conversations and keep pushing toward an objective when a task becomes difficult, rather than treating each exchange as a fresh start.
Muse Wonder is rolling out through Meta’s API and in Muse Code. The base model is priced at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is available through Muse Code’s $50-per-month power-usage tier and $200-per-month maximum-usage tier.
What Meta says Wonder can do
Meta says its Wonder models are especially capable in legal, financial and health-related work, while also providing what it calls frontier-level coding performance. Those are substantial claims, and they should be read as the company’s assessment of its own models rather than an independent verdict on their real-world reliability.
The fields listed by Meta are also among the most consequential places to position an AI system. Legal, financial and health questions can involve material consequences when a detail is misunderstood, outdated or wrong. A model may assist with organization, explanations, drafting and research-oriented workflows, but the announcement does not establish that users can safely treat its responses as professional legal, financial or medical advice. Strong model capability and appropriate use are separate questions.
For coding, the key word is performance, not autonomy. AI coding systems can be used to generate snippets, explain unfamiliar code, propose changes or help work through a programming task. But an output that looks plausible is not necessarily correct, secure or suitable for a particular project. Any developer considering an AI-generated change still needs to understand the result, test it and assess whether it fits the software’s requirements.
That caution applies especially to work that persists over time. The more an agent remembers and the longer it works toward a goal, the more important it becomes to make sure its retained context is accurate and its next action still reflects the user’s intent.
Active memory: continuity, but with a purpose
Meta defines active memory as the ability to carry what matters to a person across conversations and retrieve details relevant to the work at hand. In plain language, the intended improvement is not merely a larger conversation window. It is selective continuity: an agent should be able to remember an ongoing project, prior constraints or a user preference, then bring that information into a later task where it is useful.
That is potentially more practical than asking users to repeatedly reintroduce the same background. A person working through a multi-stage assignment may otherwise have to restate the goal, the required format, previous decisions and boundaries every time a conversation restarts. Active memory is intended to reduce that setup burden.
The difficult part is the word relevant. A useful system needs to recall the details that advance the current task without surfacing irrelevant information or acting as if an old preference applies forever. Meta’s description presents active memory as a solution to this context problem, but the announcement does not spell out what controls subscribers will have over what is remembered, how it is reviewed or how it is removed. Those questions will be important to people deciding how much personal or work information to share with a long-running agent.
Active focus: the push toward agent-style work
Meta says active focus lets an agent continue working toward a goal when the assignment gets hard. This points toward an agent model of AI use. An agent is generally understood as software aimed at pursuing a user’s objective through a sequence of steps, rather than only producing a one-off response.
In practice, an agent-style task might involve breaking a broad objective into subproblems, tracking what has already been attempted, revising an approach when it hits an obstacle and returning to the original objective. The supplied details do not say exactly which actions Muse can take, how long it can operate, or what human approval is required along the way. Still, Meta’s focus on persistence signals that it sees prolonged task completion as a key measure of progress.
This is where active focus and active memory complement one another. Memory is about carrying useful context forward. Focus is about using that context to avoid losing the thread of the assignment. The first is a question of information retention and retrieval; the second is a question of task persistence. Neither label alone guarantees correct results, but together they describe an AI assistant designed for work that cannot be finished in a single answer.
Why the pricing is worth reading closely
The Muse Wonder API price is presented in tokens. A token is a small unit of text processed by a language model; it is not necessarily equivalent to a word. Input tokens are the material sent to the model, such as a prompt, code file or conversation context. Output tokens are the text the model produces in reply.
Meta lists Muse Wonder at $4 per million input tokens and $20 per million output tokens. The substantially higher output rate means lengthy generated responses can be the larger part of a usage bill. This matters for organizations building a product on top of the API, particularly if that product asks the model to produce large explanations, extensive code or repeated analysis.
For individuals using Muse Code, the Full model is instead tied to subscriptions: $50 per month for power usage and $200 per month for maximum usage. The announcement identifies the plans but does not define the precise usage allowances, throttling rules or other limitations behind those labels. Prospective subscribers will need those particulars before they can compare the tiers in a meaningful way.
The split also suggests two distinct audiences. API pricing is geared toward developers and businesses that want to integrate the model into their own software. Muse Code plans are geared toward people who want a ready-made coding environment. AI-assisted coding is increasingly relevant to the broader software ecosystem that supports games and PC tools, where continuing updates can involve a great deal of iterative engineering. For an example of the update work players actually see, follow the ongoing coverage of Wild Woods’ 0.6.0 update.
A costly attempt to reach the AI frontier
The Wonder launch arrives after more than a year of rebuilding Meta’s AI efforts and spending tens of billions of dollars on that work. Meta has said it expects to spend as much as $145 billion this year to support its AI plans. Chief executive Mark Zuckerberg has characterized the result as requiring extensive compute investment alongside rigorous scientific work by the company’s superintelligence team.
Compute refers to the processing resources used to train and run AI models. In this context, it is a central part of the story rather than a decorative technical term. Large models require substantial computing capacity, and Meta is explicitly connecting Wonder’s development to major financial and infrastructure investment.
Meta’s assertion that it has caught up with the frontier is a competitive claim. “Frontier” AI typically means the most advanced general-purpose models and the systems that define the current upper edge of model capability. The announcement gives Meta’s position, but it does not provide a shared benchmark table or third-party evaluation that would settle comparisons with other companies’ systems. As a result, the more concrete news for users is the product rollout: which models exist, where they can be accessed, what Meta says they are designed to do and what they cost.
What users should take from the announcement
Wonder’s most meaningful proposed change is its effort to make AI less forgetful and less easily derailed during complex work. If Meta’s active memory and active focus operate as described, Muse could feel less like a chat window that needs constant re-briefing and more like an assistant with a continuing understanding of a project.
But persistent context raises practical discipline questions. Users should be deliberate about what they enter into an AI service, especially when the work involves sensitive personal, legal, financial or health details. They should also check consequential outputs rather than assuming a system’s apparent continuity is the same as verified understanding.
For developers, Wonder’s API cost structure makes input and output volume a design consideration. For Muse Code subscribers, the plan prices make it important to establish whether their workload actually benefits from the Full model. And for everyone, Meta’s claims about legal, financial, health and coding strengths are best treated as reasons to examine the product—not as a substitute for validation in high-stakes tasks.
Muse Wonder and Muse Wonder Full are now rolling out to subscribers. Meta’s spending and its language around personal superintelligence make clear that the company sees this as a major strategic marker. Whether the models meet the larger frontier claim will depend not only on impressive demonstrations, but on how reliably memory, focus, coding assistance and safety hold up in sustained everyday use.
