Meta's Muse Wonder: the API price, the data modes and the model held back from the API

Meta announced the two models in its newsroom on 12 October 2026, calling them its most capable models to date and describing them as built for personal and professional agentic tasks in domains including law, finance and health, as well as coding. The newsroom post says Muse Wonder is available to developers through the Meta Model API and Muse Code, that Muse Wonder Full is available to Muse Code subscribers, and that Muse Wonder will soon power experiences in Muse, rolling out first to subscribers.

Meta's pricing page lists Muse Wonder under the model id muse-wonder-1.0 at $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens, in US dollars. It says there is no long-context premium, so the rate is the same whether the context window is nearly empty or nearly full. Web search grounding is charged at $2.50 per 1,000 search queries on top of the token cost. The same page says pay-as-you-go carries no minimums or upfront commitment.

The models page gives a context window of 1,048,576 tokens, with text, image, video, audio and PDF input and text output, and says Muse Wonder is in preview. Meta's developer post is more careful on one input: it says audio-dependent use cases require confirmation of audio support on the release surface.

Muse Wonder Full is described on the models page as the higher-capability variant and is marked not available on the Model API. It is available only in Muse Code with a subscription, on the Power Usage and Maximum Usage plans. The subscriptions page lists four plans, Everyday Usage, High Usage, Power Usage and Maximum Usage, with High at 5 times, Power at 20 times and Maximum at 80 times the usage of the Everyday plan. The developer post says Maximum Usage is a new plan with the same five-hour and weekly resets.

Rate limits for muse-wonder-1.0 are 3,000 requests and 4,000,000 tokens per minute, applied per team and not per API key.

The price sits well above Meta's own smaller model. By our arithmetic from Meta's pricing page, Muse Wonder costs 3.2 times as much as Muse Spark on input ($4.00 against $1.25) and about 4.7 times as much on output ($20.00 against $4.25). Meta's own guidance on the models page is to use Muse Wonder where capability matters more than cost per token and Muse Spark for high-volume or cost-sensitive work. A workload that sends everything to the larger model pays for capability it may not need.

Cached input is where the bill can move most. At $0.20 against $4.00, a cached token costs 5% of a fresh one by our arithmetic, so an agent that reuses a long, stable prefix costs far less than one that rebuilds its context on every call. That is a design decision for whoever builds the agent, and it is worth making before a pilot's costs are used to forecast production.

The split between the two models is a procurement point. The variant Meta describes as its most advanced cannot be called from an application at any price published today. It is reachable only through Meta's own coding tool, on a flat monthly subscription with usage limits that reset. An organisation evaluating Meta for a product or an internal system is evaluating Muse Wonder, and should make sure any benchmark it is shown is for that model and not for Muse Wonder Full. Meta's developer post says Muse Wonder Full tops two coding benchmarks in independent third-party runs; we have not reproduced its figures here.

Preview status matters for anything with a service level. The pages we read give no date for general availability and no statement of what may change before it.

Meta's pricing page describes two modes. In Standard mode you pay standard prices and your content is not used to train Meta models. In Contributor mode you get discounted pricing or more usage in exchange for permission to use your content, which the page defines as the prompts, code, files and context you send and the responses you get back, to train future Meta models. The page states that Contributor mode is not available for Muse Wonder on the Model API.

The subscription is different. The subscriptions page says Muse Code subscriptions default to Contributor mode, that it is a setting which can be turned off at any time, and that from that point future content is not used for training. Meta's geographic use policy, last updated 12 October 2026, puts it in contract terms: anyone accessing the services through a coding subscription is considered to be using the Contributor Services by default, whichever model they select, unless they change the setting in their account. Both Muse Wonder models are available in either mode on the plans that include them.

Zero data retention is offered for qualified Model API accounts. Meta's help page says that with it enabled, prompts and responses are processed in real time and not stored after the response is returned, except where retention is required by law or to address misuse. It cannot be turned on from the web interface: it is enabled at organisation level by Meta after a request through the sales form, and the page says support may vary by model. With it on, stateful requests that reference an earlier response, background requests, file uploads, web search and playground chat history are unavailable.

The same page lists what Meta still keeps under zero data retention: the inputs and outputs of requests flagged for an Acceptable Use Policy violation, for up to two years; token, request and tool-call counts; account and team details; operational metadata such as model name, latency and app id; and anything sent through support. Calls routed through third-party platforms follow that platform's retention policy.

On geography, the policy says the services are not available in all jurisdictions, that Meta may enable or restrict access in any jurisdiction at its discretion, and it lists restricted territories. The UK, the US and the Gulf Cooperation Council states are not on that list. The policy also says any published list of available countries is for convenience and does not guarantee continued access.

The default is the risk. A developer who subscribes to Muse Code to try the stronger model is, unless they change a setting, giving Meta permission to train on the code, files and context they send. In a company setting that can be source code and whatever sits beside it in the repository. Turning the setting off stops future use; the subscriptions page says nothing about content already sent. A subscription is taken out with a Meta Account, so an individual can do this without the organisation's procurement or security team seeing it.

The API is the cleaner route for company work, because Muse Wonder there runs in Standard mode only. An organisation that wants its engineers to try Meta's models can point them at pay-as-you-go keys under a company account and make that the stated policy.

Zero data retention is not a switch a regulated buyer can assume. It needs Meta's approval, it may not cover every model, and it removes server-side conversation state, background runs, file uploads and web search, which are features an agent design may depend on. A team that needs it should confirm it for muse-wonder-1.0 and test the agent with those features off before committing to a design.

The pages we read say nothing about where data is processed or stored. For buyers with UK, EU or Gulf residency requirements that is an open question to put to Meta in writing, not an assumption to make in either direction.