Muse Wonder: What It Is, Pricing, and How to Get It

What Is Muse Wonder?

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026. The release adds a higher-capability option for work that involves planning, using tools and recovering from failed attempts over an extended task. An agent is the surrounding software that lets a model take those actions, rather than only return a chat answer. Meta's announcement dates the release and identifies the two variants.

The practical change is another model choice inside your workflow. A builder might evaluate Wonder on a refactor spanning several packages. An operator might evaluate it on a document-heavy investigation whose answer requires several tools. Neither needs to replace a working, inexpensive classifier just because a new model launched.

Meta describes planning, parallel tool use and error recovery in its developer release. Those capabilities matter only if the resulting work passes your checks. A longer run that finishes the wrong task still creates review work.

Muse Wonder vs Wonder Full: The Access Difference

Wonder is the API option; Wonder Full is the higher-capability variant restricted to eligible Muse Code subscriptions. Both are listed with 1M context on Meta's model page. Context is the material a model can consider together, measured in tokens, which are small units of text and other encoded input.

The access restrictions come from Meta's model availability documentation. That page explicitly says setting Wonder's reasoning effort to max does not select Wonder Full. Reasoning effort changes how much thinking the selected model does; it does not change your model entitlement.

What Does Wonder Full Cost?

Meta lists Muse Code Power Usage at $50.00/month and Maximum Usage at $200.00/month on the Wonder model page. Both include Full. These are Muse Code prices, not prices for the separate Muse app.

Buy Power Usage to evaluate Full if its included usage fits your work. Consider Maximum Usage when you need its larger allowance, not because Full requires the most expensive plan. Meta lists the allowances as 20x and 80x Everyday Usage, respectively; neither means unlimited work.

A subscription also does not fund arbitrary API traffic. The subscription documentation says its credential is for Muse Code only; additional API keys are billed separately. Benefits and availability can vary by region.

Is Wonder in the Muse App Yet?

Meta's launch wording is “soon,” with subscribers first. It does not establish a launch date or a Wonder Full entitlement for the Muse app. Use the API or Muse Code for a documented path today, and treat the app rollout as pending under Meta's announcement.

Muse Wonder API Pricing: What a Job Costs

Budget for the tokens a job actually consumes, including its repeated calls. The 1M context window is capacity, not a flat-price request or a monthly allowance.

These USD rates were verified against Meta's live model and pricing pages on October 13, 2026. Each price is per million tokens in Standard mode:

Cached input is reused prompt content that qualifies for the lower rate. Meta's pricing documentation says there is no long-context premium. It separately charges web search at $2.50 per 1,000 queries, in addition to tokens. Wonder's API access is Standard mode only; there is no Wonder Contributor-mode API discount.

A Worked Budget for a Builder or Operator

Assume a job consumes 200,000 uncached input tokens and 20,000 billed output tokens across its calls. This is an illustrative workload, not a measured model run. Exclude search and other non-token costs, and hold usage identical:

• Spark 1.2: (0.2 × $1.25) + (0.02 × $4.25) = $0.335 per job.
• Wonder: (0.2 × $4.00) + (0.02 × $20.00) = $1.20 per job.
• Difference: $0.865 per job, or about 3.58 times the token cost.

At 1,000 jobs, that is $335 on Spark versus $1,200 on Wonder, an $865 increase. If a founder routes only the hardest 10% to Wonder and keeps the other 90% on Spark, the same illustrative workload costs $421.50. That routing example assumes the jobs can be separated by difficulty; it does not establish that either model will succeed.

Actual token counts can differ between models. Use the equal-usage calculation to plan a trial, then replace the assumption with billed usage and reviewer time from your own jobs.

When to Keep Muse Spark 1.2

Keep a validated Spark workflow when Wonder has no demonstrated benefit on that task. Muse Spark 1.2 already has 1M context and is positioned for coding and tool use on its model page. A large context window alone is therefore no reason to upgrade.

For a senior operator sorting incoming requests, extracting fields or assembling a routine digest, the useful question is whether the output meets the acceptance rule. For a solo builder fixing a contained bug, it is whether the patch and tests are correct. If Spark already clears that bar, Wonder's higher rate needs a specific justification.

There is one version distinction to keep current: Meta now recommends Muse Spark 1.3 for new Spark work, while 1.2 remains available. Its model documentation also flags incomplete audio support in 1.3 and points audio users to 1.2. Staying on the Spark family and deliberately keeping a tested 1.2 deployment are separate decisions.

Who Should Act, Wait, or Keep Their Current Model?

Builders should trial Wonder when a task repeatedly stalls across files, tools or intermediate failures. Keep the scope and acceptance checks fixed so you can see whether the model change helps. Evaluate Full inside Muse Code only when the harder coding work justifies that subscription.

Operators should keep inexpensive work on Spark and try Wonder on a bounded exception: for example, a difficult case that needs several documents reconciled into an answer with evidence. Require the evidence you would ask of a person doing that job.

Buyers and CTOs should separate the deployment decision from the model ranking. A product that needs API integration cannot use Full through the API today. An operator waiting specifically for Wonder inside the Muse app should wait for confirmed availability there, rather than buy a Muse Code plan for the wrong product.

Active Memory and Active Focus: Meta's Claims

Meta's promise is relevant recall plus sustained work, not simply a bigger prompt. The distinction matters because remembering an instruction and following through on it are different failure modes.

Active Memory, as Meta describes it, carries useful context across conversations and retrieves what matters for the current task. Active Focus, in Meta's description, keeps working through difficult goals, tries alternatives, handles blockers and seeks clarification or checks before important actions. These are Meta's claimed model behaviors, not independently verified guarantees from this article.

For an operator, a useful memory test is whether an agent brings back a previously supplied delivery constraint when reviewing a new order. A useful focus test is whether it recovers from a failed lookup while preserving that constraint. The first checks recall; the second checks follow-through. Neither is proved by a polished final paragraph.

In a custom integration, decide where past information is stored and how it reaches the model. Do not assume that changing an API model ID automatically creates a durable memory system. Muse Code, for example, documents Markdown memory with separate scopes, showing that the surrounding product also has a role.

Try It Today: Muse Code or the API

Use Muse Code for a repository task, or the API for an application you control. A first trial should answer one question with a result you can inspect.

Start with Muse Code

Muse Code is Meta's terminal coding agent. Follow the official quickstart on macOS or Linux:

curl -fsSL https://dev.meta.ai/install.sh | sh
muse

Run it from the repository you want to work on. On first launch, use browser sign-in or an API key. Open /models and select Wonder; select Full only if your subscription includes it.

Give it a contained objective such as: “Diagnose the failing pagination test, propose the smallest fix, and show the focused test result.” Inspect the diff and evidence before accepting the change. Our how to use Muse Code guide covers the fuller installation, planning and review workflow.

Make a Direct API Request

Create an API key through the Meta Model API dashboard's API keys → Create API key flow, and set it in your shell as MODEL_API_KEY. With that variable set, this adapts Meta's documented Responses request to Wonder's model ID:

curl -sS https://api.meta.ai/v1/responses \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "muse-wonder-1.0",
    "input": "Plan a contained fix for a pagination bug. List the evidence needed to verify it."
  }'

The response is a plan, not a repository edit: this request supplies no repository or tools. Meta says Wonder and Spark accept the same requests on the same endpoints, making an existing Spark integration a practical starting point for evaluation.

What's Overhyped: Access Is Not Autonomous Completion

The most consequential mistake is buying access under the wrong assumption. Keep these distinctions in the purchase decision:

• Full is a separate variant. Wonder at max reasoning is still Wonder. The $4/$20 API rates do not buy Full.
• 1M context is not guaranteed recall. Capacity does not prove the model found the relevant fact or respected it throughout the task.
• A subscription is not unlimited API credit. Muse Code billing and separately created API keys have different meters.
• This is a hosted release. The cited release and access docs do not offer downloadable Wonder weights for self-hosting.

There is also a concrete setup choice: Muse Code subscriptions default to Contributor mode, which allows Meta to use supplied content for model training in exchange for more usage. Turn it off in settings if that is unsuitable for your work. In Wonder's API Standard mode, Meta says content is not used to train its models. Check the surface you actually use.

For the separate model-versus-model decision, read Muse Wonder vs Claude Opus 5.5. For background on the earlier Spark generation, the Muse Spark 1.1 review provides historical context; use the current rates above for today's budget.

Frequently Asked Questions

Your Next Move

Pick 10 contained tasks from an existing workflow. Keep the same inputs and acceptance rules, compare your current Spark setup with Wonder, and record accepted results, total billed usage and reviewer time. Keep Spark on the work it already handles; adopt Wonder where the measured improvement covers its extra cost. Evaluate Full separately only if Muse Code is where that work belongs.

For practical guides to AI model selection and operating costs, join the newsletter.