Muse review: App, Spark and Wonder Explained
Muse is worth a bounded pilot when your task crosses connected apps. Separate the consumer agent, model access and subscription entitlement; an announcement about Wonder does not prove that your current app session uses it.
Source checked October 19, 2026. Based on published product and research sources.
What changed in October
Meta’s October 12 announcement, updated October 13 introduced Muse Wonder and Wonder Full. It says Wonder is available through the Meta model API and Muse Code, with Wonder Full for Muse Code subscribers. It describes Wonder reaching the Muse app soon, starting with subscribers. That is a rollout statement, not confirmation that every Muse app session already uses Wonder.
Three things to keep separate
- Muse app: a personal agent that works with connected services.
- Muse Spark: an earlier model family; its existing profile remains a separate historical product reference.
- Muse Wonder: the newly announced model family. Earlier Spark token prices do not establish Wonder pricing or the Muse app subscription price.
Availability and business use
The small-business announcement, updated October 8 identifies US and Canada access and connectors including Notion, Slack, Shopify and Stripe. Meta says publishing, sending and spending require approval. These are vendor-described controls; this desk has not tested a connected account or an approval flow. Connector availability should be checked in app settings.
Who should evaluate it
A small business with repeated tasks across connected apps can start with a bounded draft or analysis task and inspect every proposed action. Evaluate whether the agent can identify the right records, explain its work and recover from a failed step. Access to a model API alone does not reproduce the app’s connectors or permission system.
What this review establishes
We checked the official rollout and product distinctions. We have no first-hand completion rate, safety audit or current checkout quote, so the earlier numerical score has been removed. Confirm region, plan, connector permissions and usage caps before committing a business workflow.
Related: Muse Spark profile · Spark pricing reference (not Wonder pricing) · Higgsfield creative suite update.
Build the right product map
| Layer | What to verify | What not to infer |
|---|---|---|
| Muse personal-agent app | Your region, connector permissions, task limits and approval controls | An API token price does not price the app. |
| Muse Code and model API | Available model IDs, subscription/API entitlement and metering | Access here does not establish the same model in every app account. |
| Muse Spark reference | Historical model capability and pricing information | Spark prices and context limits are not automatically Wonder specifications. |
| Wonder / Wonder Full rollout | The exact release channel and what the account actually exposes | “Coming soon” does not mean the rollout is complete. |
This distinction matters when comparing a personal assistant with a model vendor. A model responds to inputs through a particular interface. An agent product also needs connectors, permissions, state, retries and a way to approve actions. Better model benchmark results do not by themselves prove that the entire task succeeds.
Three tasks that make a useful Muse pilot
Read-only business summary. Connect only the source needed for a limited question, such as summarizing a selected set of store records. Check whether the answer cites the right period, handles refunds and avoids inventing missing data. A fluent summary is not a reconciliation.
Draft for approval. Ask for a campaign draft with a defined audience, approved facts and a fixed output format. Review the source of every claim and inspect any proposed publishing action. Evaluate the usefulness of the editable draft, not just the speed of generating it.
Cross-app follow-up. Use a non-sensitive sample task with an explicit stop before any external action. Confirm which records were read, what change is proposed and whether failure leaves partial work behind. A task should be judged on the completed state and recovery behavior as well as the first answer.
Evaluation record for an agent, not a chatbot
| Record | Example question |
|---|---|
| Task boundary | Which system and records may be used? What is excluded? |
| Success condition | What exact final artifact or state counts as complete? |
| Approval point | What will be shown before sending, publishing or spending? |
| Evidence | Can the output be traced to the original records? |
| Failure recovery | What happens when a connector times out or a record is missing? |
| Cost and limits | What usage meter or plan boundary changed during the task? |
These questions form a proposed pilot; we have not observed Muse completing them. Keep a task log that distinguishes an unsupported instruction, a connector failure, an incorrect answer and an action performed without the expected approval. Lumping every failure into “model quality” makes it harder to decide whether a product fits.
How to compare Muse with a manual or existing workflow
Use the same task definition and acceptance criteria. Record total time including permission setup, corrections and final verification. If a task takes two minutes to do manually but ten minutes to audit, a convincing demo may not create value. Conversely, a draft that saves an hour while remaining easy to verify can be useful without any autonomous action.
For small-business tasks, start with records you already understand and keep the first pilot narrow. A connector list is not proof that every action in each application is supported. Use the actual available permissions and task results to decide whether to expand access.
What new evidence would change this review?
An account-specific rollout confirmation would establish access, not task reliability. A saved task log with the inputs, approvals and completed result would support a workflow finding. A measured comparison across repeated tasks would support a completion-rate statement. A security audit would be separate evidence again. We will not replace one type with another or award a numerical score from a launch post.
Download the agent task evaluation sheet to track these distinctions. The file is blank and contains no claimed Muse results.
Frequently asked questions
Is Muse Wonder already in everyone’s Muse app?
The October announcement describes a phased app rollout beginning with subscribers. Check the model and entitlement exposed in your own account. Model API or Muse Code availability does not establish a completed consumer-app rollout.
Is Muse Spark being renamed to Wonder on this site?
No. They remain separate references. The Spark profile and pricing page describe that model family and should not be used to infer Wonder prices or limits. This review explains their relationship without merging the products.
Can Muse publish or spend without approval?
Meta says publishing, sending and spending require approval. This is a vendor-described control that this desk has not tested. A useful pilot should verify what is shown at the approval point and what happens if you decline.
How should I compare the subscription price with API prices?
First define the workload and interface. App subscriptions can include product features and usage rules that differ from token-metered API access. Compare the total cost of completing your task in each route, not just two unrelated headline prices.