Meta Releases Muse Wonder, Keeps Full in Muse Code
Developers can access Wonder through an API, while the Full model requires a Muse Code subscription, creating different options for building and using agents.
Listen
AI narration
10:23
0:00 / 10:23
AI SummaryGenerated from this article
Meta released Muse Wonder, accessible via API at $4 per million input tokens and $20 per million output tokens, alongside Muse Wonder Full, available only through a $50–$200 monthly Muse Code subscription. Both models claim benchmark parity with frontier competitors like GPT-6 Astra and feature Active Memory for cross-conversation context and Active Focus for persistent task completion. Developers should independently verify Meta's safety claims, memory capabilities, and benchmark results before production use, as the models serve different deployment paths and require separate evaluation.
Developers can build applications with Meta’s new Muse Wonder through the Meta Model API. Access to the more capable Muse Wonder Full, however, requires a Muse Code subscription and does not include an API endpoint.
In its October 12 announcement, Meta describes the pair as its most capable models yet, aimed at personal and professional agentic tasks. The company emphasizes coding, work across multiple steps, and two named capabilities: Active Memory, for carrying relevant information across conversations, and Active Focus, for continuing through difficult tasks and recovering from errors.
Builders evaluating the release need to distinguish those access paths, separate Meta’s memory claims from application-level implementation, and treat its reported benchmark parity with competing frontier models as a company claim, not an independently established overall ranking.
Wonder Full Requires a $50 or $200 Muse Code Plan
Meta’s developer product page gives more specific access terms than the launch announcement’s broad reference to Muse Code subscribers.
Access detail
Muse Wonder
Muse Wonder Full
Meta Model API
Available
Not offered
Muse Code
Available
Available on Power Usage and Maximum Usage plans
Listed context window
1 million tokens
1 million tokens
API input price
$4 per million tokens
No API price
API cached-input price
$0.20 per million tokens
No API price
API output price
$20 per million tokens
No API price
The Power Usage subscription costs $50 per month, while Maximum Usage costs $200 per month. Meta lists Wonder Full access under both plans, with advertised usage allowances of 20 times and 80 times the Everyday Usage plan, respectively. These are relative allowances, not published token quantities or promises of unlimited access.
For a developer building a customer-facing application, Wonder is the directly accessible choice through Meta’s API. A team can integrate it into its own orchestration system, attach tools, and measure performance inside its product.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
Wonder Full is available for work conducted through Muse Code, which Meta describes as a multi-agent coding tool in the terminal. That subscription provides access within Meta’s coding environment, without equivalent access for an independently hosted application. The difference limits how the two models can be compared in practice.
Meta also says Wonder will power more experiences in Muse, initially for subscribers. This is a separate rollout commitment; it does not mean every Muse experience already uses the new model.
API Costs Depend on How Much Work an Agent Generates
Wonder’s listed API rates place substantially more cost on generated output than on uncached input: $20 versus $4 per million tokens. Eligible cached input costs $0.20 per million tokens.
A workload consuming 1 million uncached input tokens and producing 100,000 output tokens would cost $6 at those published token rates, excluding any separate charges associated with tools or other services.
For agents, the useful unit of comparison is often the completed task. A workflow that repeatedly inspects files, invokes tools, delegates work, and revises its answer can consume tokens across many calls. A lower-priced request may still produce a more expensive completed job if it requires more attempts.
Caching could help workflows that repeatedly supply the same eligible material, such as stable instructions or repository context. Builders need to establish what their application can actually cache before applying the cached rate to recurring input.
The 1-million-token context window measures capacity, not reliable recall. Accepting a large body of information does not establish how accurately the model will retrieve a particular detail, reconcile conflicting documents, or maintain task priorities throughout that context.
Full’s subscription price cannot be converted into a meaningful API-token comparison from the published information alone. Its usage allowances and product environment differ from Wonder’s metered API access.
Active Memory and Active Focus Need Application-Level Scrutiny
Active Memory is the ability to carry relevant information across conversations and retrieve details that matter to the current task. Active Focus describes continued work toward a goal: trying alternative approaches, recovering from errors, asking for clarification, and checking before important actions.
Memory and persistence address different failure modes. A model might handle a large prompt well yet fail to preserve a user’s preferences between sessions. It might remember those preferences but abandon a task after a tool error.
The launch material leaves the implementation boundaries unclear. It does not explain how much of Active Memory depends on the model itself, a separate memory store, retrieval, or the surrounding Muse product. An API integration should not assume that separate requests automatically share persistent memory merely because the model family carries that label.
Builders also need answers about control: what gets retained, how users inspect or remove it, how outdated information is corrected, and how information stays separated between users or projects. The cited product material does not provide enough detail to settle those questions.
Active Focus describes desired behavior without guaranteeing safe execution. Continuing after an error is valuable when an agent can diagnose the problem. It is less valuable if the agent repeats an incorrect action or pursues an obsolete instruction.
Applications still need explicit permissions, stopping conditions, and approval requirements. A model’s tendency to check before something important should complement those controls, not replace them.
Meta’s Frontier-Parity Claim Comes From Its Own Evaluations
TechCrunch’s launch report describes the release as Meta’s first attempt at a frontier model and reports that Meta’s testing places Wonder and Wonder Full at comparable benchmark levels to Fable 5.1 and GPT-6 Astra.
These are results presented by Meta. They do not establish that the models are interchangeable across tasks.
Meta’s developer page includes evaluations spanning coding, computer use, professional work, and memory, along with an internal composite called the Personal Superintelligence Index. The range fits the company’s agent positioning, though its internal aggregate is not an independent measure of overall model quality.
Meta also says Wonder Full leads DeepSWE and SWE-Marathon in independent third-party runs. That claim is more specific than a general assertion of frontier performance, but it appears on Meta’s product page. Drawing a firm conclusion about the broader comparison still requires examining the underlying runs, model settings, tool environments, and resource budgets.
The published comparison labels model configurations by effort settings, including “max.” A headline score may therefore not represent the latency and cost of a builder’s intended production configuration.
On memory, TechCrunch reports Meta’s claim that Wonder outperforms the listed competing models on MemoryBench, which evaluates continual-learning and memory-construction frameworks. This supports Meta’s positioning without proving that every API application will inherit the same cross-session behavior.
The most useful next comparison would test the accessible Wonder API against alternatives on identical tasks, with the same tools, budgets, and completion criteria. Full’s results should remain separate unless the intended deployment runs inside Muse Code.
Deployment Controls Matter More as Agents Gain Persistence
Meta says it trained the Wonder family to avoid sharing private information without permission, refuse assistance with harmful acts, and support legitimate defensive security work. Its launch announcement also says a fuller report containing additional safety evaluations will be available soon.
At launch, then, the announcement explains intended safeguards without providing the full evaluation record needed to assess them.
Meta explicitly envisions agents working across messages, calendars, and documents. Persistent access to that material raises questions beyond answer quality: which information can reach a tool, which actions need confirmation, and what happens when a task continues after circumstances change.
A defensible initial Wonder evaluation would measure completed-task cost, recovery from deliberately introduced tool failures, and compliance with approval boundaries. Memory testing should cover incorrect retention and cross-project leakage as well as successful recall.
The access split creates two separate purchasing decisions. A developer can evaluate Wonder as infrastructure for an application; a coding team can evaluate Wonder Full as part of a Muse Code subscription. A strong result in the latter environment does not automatically answer the former question.
Whether Wonder delivers reliable agent performance at a competitive completed-task cost, and whether Full’s advantages justify working inside Meta’s coding product, remains to be tested independently. Until then, access conditions and deployment behavior deserve as much attention as the benchmark charts.
