Meta launched Muse Wonder and Muse Wonder Full on October 12, introducing two AI models intended for work involving applications, documents and tasks that unfold across multiple steps. Muse Wonder is available through Meta Model API, where the documentation labels it as a preview, and in the Muse Code coding agent. The Full variant has a different distribution route: access is provided through selected Muse Code subscriptions, rather than the pay-as-you-go API.
The distinction is explicit in Meta's technical catalog. Muse Wonder Full is included in the Power Usage and Maximum Usage plans. Setting the API version of Muse Wonder to max reasoning effort does not automatically select Full. The two names identify variants with separate access conditions, rather than two labels for the same request setting.
API access and Muse Code subscriptions
Published Standard-mode rates for Muse Wonder are $4 per million input tokens and $20 per million output tokens. Cached input costs $0.20 per million tokens. These are usage-based prices in US dollars, separate from a monthly Muse Code subscription and any additional services used within an application.
The Muse Code page lists Power Usage at $50 per month and Maximum Usage at $200 per month. They offer 20 and 80 times the usage of the Everyday Usage plan, respectively. Everyday and High plans include Muse Wonder but not the Full access reserved for higher tiers. For the new Maximum plan, Meta specifies that the five-hour and weekly reset windows remain in place.
The API model is identified as muse-wonder-1.0. Its documentation describes a context window of 1,048,576 tokens and inputs that can include text, images, video, audio and PDF documents. The model returns text. This combination does not mean Muse Wonder automatically generates photographs, video clips or speech: Meta provides other models and product surfaces for those outputs.
Memory and continuity for longer tasks
According to Meta, Active Memory is designed to carry important context across conversations and retrieve relevant details when they become necessary. The company describes this capability in terms of goals, constraints and earlier decisions. It is presented as part of the agent's behavior, rather than a claim that the model already knows a company's information or has default access to its documents and applications.
Active Focus concerns continued work when a task becomes difficult. Meta says the models are trained to try alternative approaches, adapt to new information and recover from errors. That behavior also includes asking for clarification and checking the user's intent before important actions. The description is not a guarantee that every task will finish correctly or without human intervention.
For coding, the company highlights changes spanning multiple files, debugging, testing and successive project revisions. Meta's published evaluations include extended tasks, such as SWE-Marathon, and tests of sustained focus. The results are reported as model evaluations under particular configurations; they are not measurements of the productivity of every team that will use the product.
The model and application tools have different roles
Meta Model API allows the model to be used in custom applications with AI integration, while Muse Code provides an agent ready for terminal work and development workflows. The documentation separates calling the model from using the complete agent. For developers, Muse Wonder and Muse Spark accept the same request formats on the documented API surfaces, while differing in capabilities and cost.
The technical catalog lists tool calling, structured output, image understanding and web search among Muse Wonder's capabilities. Meta's examples treat a tool call as one stage in a loop: the application detects the request, performs the operation and sends its result back to the model. Analyzing a requirement and actually carrying out an action in a system are therefore separate operations in the integration described.
Muse Code separately documents approvals, command isolation and records of agent actions. Its recipe catalog includes session recovery, staged command approvals and rules that the agent cannot rewrite without review. Those mechanisms belong to the environment in which the agent runs. They are not equivalent to simply selecting a model name within a third-party application.
Different conditions for using submitted data
In Standard mode, Meta says submitted information and generated responses are not used to train its models. Contributor mode has different conditions: it offers more usage or lower costs in exchange for permission to use content to train future models. For Muse Wonder, Contributor is unavailable through the pay-as-you-go API, although it exists for Muse Code subscribers.
Contributor is enabled by default in Muse Code subscriptions and can be turned off in settings. The documentation includes prompts, code, files, context and model responses in the covered content. Subscription conditions should therefore be distinguished from those of Standard mode on the API. The pricing page also specifies that API rate limits apply to a team, rather than separately to each access key.
Meta says the Wonder series was trained to protect private information, resist adversarial instructions and limit harmful uses while retaining certain defensive cybersecurity applications. These are the company's statements about product behavior and evaluation. At launch, Meta also announced an additional safety report that would be published later.
Muse Wonder's integration into experiences in the Muse app is planned for the coming period, beginning with subscribers. That stage is not presented as access already being active for every user of Meta's social products. The documentation warns that model availability can vary by country; the API launch, access through Muse Code and distribution within the consumer app remain three situations that must be checked separately.
