You see a result for Muse Wonder Full, open the API documentation, set reasoning effort to max and expect the same model. That is an easy mistake to make after this launch. Meta announced Muse Wonder and Muse Wonder Full on October 12, but their access paths are different.
TechCrunch and Quartz have covered the announcement. For application builders, the useful questions are which version they can access, what a completed task costs and which data mode applies to submitted code and files. This article draws on current official documentation. Cost examples are calculations, not measured bills.
Check the version: max does not turn Wonder into Full
Meta’s model documentation is explicit: Muse Wonder is available through Meta Model API and Muse Code. Muse Wonder Full is currently offered only through the Power Usage and Maximum Usage subscriptions in Muse Code. Setting ordinary Wonder’s reasoning effort to max does not select Full.
API developers should check the actual model ID instead of inferring what they are calling from a product name or benchmark screenshot. The documented standard model ID is muse-wonder-1.0. Full’s results and access should be discussed separately, rather than treated as a promise for the API model.
Individual users should also distinguish a model launch from an application update. Meta says Wonder will subsequently power Muse experiences, rolling out to subscribers first. This does not establish that every Muse user already has free access. Check the current entry point for regions, account eligibility and plan benefits.
Why long tasks still need a budget
Checked on October 13, the official standard API rates are $4 per million input tokens and $20 per million output tokens. Input and output are calculated separately. These are not monthly fees or a $4 price per task.
Consider an example counting only uncached text usage. A request containing 10,000 input tokens and producing 2,000 output tokens costs $0.04 for input and $0.04 for output, or $0.08 in total. Ten requests with that same usage would total $0.80 in text charges.
Ten requests is an illustrative assumption, not a claim about a typical task. Caching, search tools, retries and other charges are excluded. Actual cost also depends on how much history each round includes and how much output it generates. The official page separately lists search-tool charges, which should also be checked when integrating.
AI Islands recommends tracking the total usage needed to complete a piece of work. For code changes, compare whether the change passes tests, how much human correction it needs and how often a failed attempt is rerun. The price of one response alone is a poor guide to which model is economical for your task.
Extra subscription usage also involves a data mode
Meta distinguishes Standard and Contributor modes. Standard does not use submitted content to train Meta models. Contributor permits that content to be used for future training in exchange for discounts or more usage. Muse Code subscriptions default to Contributor, which can be disabled in settings. Wonder on Model API currently does not offer Contributor mode.
Do not assume that every entry point follows the same rule. Before a trial, check the actual product and current setting, then decide which code, files and context to submit. Standard’s training-use statement should not be expanded into a claim of no data retention. Retention and other handling terms require separate verification.
How would you test memory and persistence?
Meta describes Active Memory as retaining and retrieving important context across conversations, and Active Focus as continuing through obstacles, asking for clarification and checking important actions when needed. These are the company’s descriptions of the design. This article does not present independent testing.
A small evaluation could ask the model to change a form in a test repository, specify files it must not touch and require test results. Add a constraint in a second round and check whether it preserves the earlier requirements. This is an editorial test suggestion, not an established performance result.
The question is whether continued attempts respect boundaries. Specify what it should report when permissions, a budget or evidence are insufficient. The Agent Skill usage and safety notes can help you decide which tools to make available.
When is a trial useful, and when can a switch wait?
If you regularly delegate multi-step code changes, tool calls and work across conversations, Wonder is a candidate for a limited comparison. Keep the existing workflow, run the same tasks and compare delivery quality, total cost and rework before deciding whether to migrate.
If short text organization is your main need and the current tool is reliable, a new launch need not trigger a rebuilt workflow. Establish the version, cost and data mode, then look for benefit in your own tasks. The AI tool directory can support everyday selection; an untested model should not be described as a universal replacement for existing products.