Does Meta Train on Muse Code? Contributor Mode, Opt-Out and Wonder Access

Yes. Meta can use content submitted through a Muse Code subscription to train future models while Contributor mode is enabled. That mode is on by default. You can turn it off in /settings, but your usage allowance changes and the opt-out applies to future submissions.

Choosing a different model does not opt a subscription out. The account setting governs how its content is treated, including when using Wonder or Wonder Full.

Direct API access follows a different arrangement. Meta lists regular Wonder as Standard mode only, with submitted content excluded from model training.

That gives developers two decisions to make: which model fits the work, and which data arrangement fits the information being submitted. The first does not settle the second.

What Contributor mode covers

Meta’s Contributor-mode explanation includes prompts, code, files, context and model responses.

Its terms also include content the service is authorized to access through connectors, integrations or background context.

The policy concerns material submitted to the service. It does not establish that Meta automatically receives every file in every repository. However, a repository being private does not change the rules for content an agent reads and sends while completing a task.

A coding request can therefore contain considerably more than the sentence typed into the prompt box. Project files, instructions and tool results may contribute to the request.

That matters for client work. A public demonstration and a confidential application can use the same agent while requiring different data settings.

Picking a model does not opt a subscription out

Some older Muse Code guides describe changing between model IDs with and without a -contributor suffix.

That distinction applies to the API model variants Meta offers. Under a Muse Code subscription, the account setting determines Contributor or Standard treatment regardless of the model selected.

Switching models inside the subscription is therefore not a substitute for checking /settings.

Meta’s Help Center makes the distinction explicit: Contributor mode is a subscription setting in Muse Code, while API users select a Contributor variant where one is available.

For regular Wonder, Meta currently offers no Contributor variant on Model API.

How to turn Contributor mode off

Meta documents the following procedure:

1. 1Open /settings in Muse Code.
2. 2Turn Contributor mode off.
3. 3Check the resulting allowance before relying on the plan for recurring work.

The plan stays the same. Meta says the amount of included usage changes because Contributor mode provides more usage in exchange for training permission.

The reviewed documentation does not publish a numerical comparison of Standard and Contributor allowances. It also establishes no separate opt-out surcharge.

That leaves an important buying question unanswered: how much useful work fits the subscription after choosing the required data setting?

Advertised multipliers cannot answer it alone. A short edit, a large repository task and a lengthy reasoning session are different workloads.

Check the setting before submitting private material.

Opting out is not retroactive

Meta’s terms make the election prospective. Turning Contributor mode off does not withdraw its existing rights over content previously submitted through Contributor Services.

The terms also state that previously Standard content becomes Contributor content if included in a later Contributor request, directly or as context.

Meta prohibits submitting sensitive, confidential or personal information to Contributor Services. It specifically includes code that users intend, or are contractually required, to keep confidential.

For developers handling proprietary work, the data setting belongs at the beginning of the workflow. Changing it after submission does not undo the earlier permission.

Standard mode does not mean zero retention

Standard mode excludes submitted content from model training. It does not establish that every prompt, response and operational record is immediately deleted.

Meta’s terms allow retention for service delivery, legal compliance, security, safety, abuse handling and other stated purposes. The reviewed terms do not establish one universal deletion period for ordinary use.

Zero Data Retention is a separate arrangement for qualified Meta Model API organizations. Meta must enable it after an eligibility request; it is not an ordinary web-interface setting.

ZDR support can vary by model. It also disables features including server-side stateful requests, background requests, file uploads, web search and playground chat history.

Contributor mode is unavailable to ZDR accounts. Legal and misuse exceptions remain, including retention of flagged inputs and outputs for up to two years. Third-party integrations have separate policies.

Teams need separate answers about training permission, retention and the features their workflow requires.

Wonder, Wonder Full, Muse Code and Muse are different products

Meta’s model documentation confirms regular Wonder on every Muse Code plan. Full requires Power or Maximum.

It also explicitly states that setting regular Wonder’s reasoning effort to max does not select Full.

Full’s subscription access does not create an API entitlement. Meta lists it as unavailable on Model API, and subscription credentials are restricted to Muse Code.

For developers integrating a model into an application, regular Wonder is the available direct-API variant.

How much does Muse Code cost?

Meta’s current subscription Help Center lists four plans.

The multipliers are relative to Everyday. They do not promise a fixed number of maximum-effort Full tasks.

Meta’s developer launch post describes Maximum as having the same five-hour and weekly resets.

Meta also says subscription benefits and availability can vary by region. Check the benefits offered to your account before paying.

The subscription documentation draws another boundary: additional API keys under the same account incur pay-as-you-go charges. A Power subscription does not fund API traffic from your own application.

Before comparing Muse Code with Claude Code or Codex, select the data mode you require and assess allowance consumption on representative work.

What does Wonder’s API cost?

Meta lists muse-wonder-1.0 at:

• $4 per million uncached input tokens.
• $0.20 per million cached input tokens.
• $20 per million output tokens.

Its pricing documentation specifies no long-context premium. Built-in web search costs an additional $2.50 per 1,000 queries, plus token charges.

Published Wonder limits are 3,000 requests and four million tokens per minute per team. Creating more API keys does not create independent team allowances.

For illustration, 50,000 uncached input tokens and 10,000 billable output tokens cost $0.40 at those rates, before search and other applicable charges. This is a calculation, not a measured task cost.

The API overview identifies Wonder as a preview model with a 1,048,576-token context window.

A large context window is a capacity specification. It does not establish perfect recall or economical completion of every large task.

Is Wonder’s API excluded from training?

Meta currently lists Wonder’s direct API access as Standard mode only. Its pricing page explicitly says Contributor mode is unavailable for Wonder on Model API.

Meta’s model designation policy, updated October 12, also lists Wonder and Wonder Full among its Standard Models.

The same policy explains that a coding subscription defaults to Contributor treatment regardless of the model selected. Users must change the account setting to use Standard treatment through that subscription.

The apparent contradiction disappears when the access arrangement is included:

The model name alone is insufficient to determine the subscription’s data handling.

Is Wonder cheaper than Claude or OpenAI?

Wonder’s standard input and output prices match Claude Opus 5.5: $4 and $20 per million tokens.

Original GPT-6 Sol lists $2 input and $10 output at its base Standard rates. Above 272,000 input tokens, its input and cache rates double and its output rate increases by 50% for the full request.

GPT-6.1 Sol is a separate, newer model.

Wonder therefore has no automatic unit-price advantage over Opus or original Sol. A stronger result on a particular task could still justify its cost, but that needs evidence from the workload.

Repeated context, reasoning, retries and human correction affect the cost of accepted work. Subscription allowances add another constraint.

A token-price table builds a shortlist. It does not determine which agent earns its place.

What do the benchmarks establish?

Artificial Analysis lists regular Wonder at max effort at 52 on its Intelligence Index and Wonder Full at max effort at 56.

Its retrieved leaderboard lists Claude Opus 5.5 at max effort with default fallback at 58.

Those results support taking Meta’s models seriously. They do not establish a universal winner or allow regular Wonder to inherit Full’s performance.

Meta’s launch scorecard also reports 85.6 for regular Wonder on MemoryBench. That is a named memory evaluation, not an overall intelligence percentage.

Coding comparisons require the model, reasoning setting, agent harness and task definition. Results for older Muse Spark configurations cannot be relabeled as Wonder results.

For broader model selection, see Choosely’s GPT-6 Sol versus Claude Opus 5.5 comparison.

Can existing OpenAI or Anthropic clients use Wonder?

Meta documents OpenAI-compatible Responses and Chat Completions interfaces, plus an Anthropic-compatible Messages interface.

OpenAI SDK clients use https://api.meta.ai/v1. Anthropic SDK clients use the base host https://api.meta.ai. Credentials and the model ID must also change.

Compatibility has boundaries. Chat Completions rejects parameters including stop, n greater than one, verbosity, logit_bias, prediction, web_search_options, modalities and audio.

Built-in web search runs through Responses. Chat Completions does not carry reasoning between turns.

A familiar SDK reduces integration work. It does not establish identical request validation, tool behavior or reasoning continuity.

Sources: API formats and Chat Completions compatibility.

Is Wonder available in the Muse app?

Meta says Wonder will soon power experiences in Muse, starting with subscribers. It has not provided a rollout date or confirmed a universal free entitlement.

Availability and subscription benefits can vary by country and account. Check the benefits shown during onboarding before subscribing.

Consumer app access, Muse Code subscriptions and API eligibility are separate questions.

Choosely’s verdict

Muse Wonder earns a controlled evaluation. Wonder Full is a credible candidate for difficult repository work, but subscription access and usable allowance need testing against the work you actually do.

For proprietary code or confidential client material, settle the data setting first. Under a subscription, changing models does not opt out of Contributor mode. Change the account setting, then assess what the plan delivers.

Use a bounded task with clear acceptance criteria. Compare passing tests, useful output, intervention time and allowance consumed. Keep an established workflow until the replacement earns its place.

Meta has become a more serious option. Its subscription default deserves the same scrutiny as its benchmark claims.

Frequently asked questions

Does Meta train on Muse Code?

It can use submitted content for training while Contributor mode is enabled. Muse Code subscriptions turn that mode on by default.

Does selecting a non-Contributor model opt a subscription out?

No. Subscription data handling follows the account setting regardless of the selected model. Turn Contributor mode off in /settings.

Can Wonder Full be used without Contributor mode?

Yes. Meta lists Full in either mode on the Power and Maximum plans. Turning Contributor mode off changes included usage.

Is Wonder Full available through an API?

Meta currently lists Full only through eligible Muse Code subscriptions. Regular Wonder is available through Model API.

Does opting out increase the subscription price?

Meta says it changes included usage without changing the plan. The reviewed documentation establishes no separate surcharge.

Does a private repository prevent Contributor training?

Repository visibility does not change the rules for content submitted to the service. Check the data setting and what the agent sends.

Does Standard mode mean prompts are immediately deleted?

No. Standard mode excludes model training. Zero Data Retention is a separate arrangement with its own eligibility and feature restrictions.

Is Muse Wonder free?

The verified developer routes are metered API usage and paid Muse Code plans. Meta has not confirmed a universal free consumer entitlement.

Is Wonder better than Claude Code or Codex?

Launch evidence does not establish a universal replacement. Compare the exact model, agent, data setting and allowance on representative work.

Track meaningful changes to the AI tools in your stack with Choosely Stack Intelligence.