Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Wonder is available through Meta Model API and Muse Code. Wonder Full is limited to eligible Muse Code Power and Maximum plans, where available. The consumer Muse app has a separate rollout, starting with subscribers.
In this guide, you will give Wonder a workshop brief, check its agenda and budget, then change the requirements. The final change makes the original plan impossible. A useful answer should explain the conflict and ask what may change.
We tested the three prompts below with standard Muse Wonder at High effort in Muse Code 1.4.6. All three ran in one retained conversation using Code's headless mode. The first answer, attendance revision, and schedule-conflict answer passed our checks. No tools or external actions ran. We did not test Wonder Full or Spark in this exercise.
University Pro members get the prepared briefs, change requests, answer key, and comparison record. The exercise also appears below. It creates no bookings, sends no messages, and requires no app build.
What is Muse Wonder?
Muse Wonder is a model available through Meta Model API. Muse Code is the terminal app we use to access it here. Muse is Meta’s personal-agent product. These names describe different parts of the system; confirm which one your account offers.
Meta describes two aims for the release in its Wonder announcement. Active Memory retrieves relevant context across conversations. Active Focus helps the model continue a difficult task, adapt to changes, and ask for input. Treat these as Meta’s capability claims.
Our short exercise checks whether a model follows and updates a brief. It does not measure persistent memory across chats or performance on hours-long tasks.
Muse Wonder vs Wonder Full vs Spark
Model
Confirmed access
Our suggested use
Muse Wonder
Meta Model API and Muse Code
A first test on work with several rules, source details, and revisions.
Muse Wonder Full
Eligible Muse Code Power and Maximum plans
Compare it on the hard cases your current model misses.
Muse Spark 1.3
Meta Model API and Muse Code
Keep it as a lower-cost comparison for routine work and existing projects.
Meta positions Wonder Full as its most capable option. Spark 1.3 remains a supported model for coding and tasks with several steps. The suggested uses above are starting points for your own comparison.
A Muse app subscription and a Muse Code subscription are different access checks. Do not buy one assuming it includes every product called Muse. Meta documents the Code subscription and its dedicated credential separately.
What does Muse Wonder cost?
Meta’s current Model API pricing lists these rates in US dollars per million tokens. Tokens are the small units a model processes.
Model
Input
Cached input
Output
Muse Wonder
$4.00
$0.20
$20.00
Muse Spark 1.3, Standard
$1.25
$0.15
$4.25
Muse Wonder Full
No public per-token rate listed
No rate listed
Eligible Power or Maximum plan
These prices apply to API usage. Check the separate Muse Code plan and its benefits in your account.
For an equal-usage example, 20,000 uncached input tokens and 5,000 billed output tokens cost $0.18 with Wonder or $0.04625 with Spark 1.3 Standard. This is an arithmetic example. The workshop exercise's actual cost depends on all turns, tools, and review.
Check your billing route before running. A Muse Code subscription applies to its Code credential; additional API keys use pay-as-you-go billing. Region and plan benefits can differ.
Meta’s published scorecard has mixed results. Wonder Full leads the listed models on DeepSWE and SWE-Marathon; Claude Opus 5.5 leads ProgramBench. The report combines internal, provider, and third-party results with method notes. Some comparisons use different agent systems. This does not establish one winner for every task.
Our exercise also cannot rank coding ability or browser use. Keep conclusions tied to what you tested.
Can you use Wonder in the normal Muse app?
The launch announcement says Wonder will reach Muse, beginning with subscribers. It does not promise immediate access for every consumer account. Check the actual product before presenting a Muse app conversation as a Wonder test.
Our Meta Muse Desktop guide covers the personal agent’s setup and permissions. Use it for that product; the Muse Code commands here belong to a different app.
Build a useful test before switching models
Choose a few jobs you already do, write down what a correct result requires, and include at least one case the current model gets wrong.
Nate Grahek’s Intro to AI Evals teaches how to use real inputs, expected answers, and repeatable checks to compare quality, speed, and cost. Its examples use a spreadsheet and Claude Code. Apply the testing method here; the recording is not a Muse Wonder lesson.
University Pro members can watch Intro to AI Evals and get this guide’s briefs, change requests, answer key, and comparison record. Meta subscriptions and usage have separate terms.
Step 1
Open Muse Code and select Wonder
The launch-day route is developer-focused, so this walkthrough uses a terminal. You will type setup commands, then give the task in plain English.
Use Meta’s official Muse Code setup instructions for your operating system. Its installer supports macOS, Linux, and Windows. Review an installer before running it; it installs software on your computer.
On macOS or Linux, download the official installer:
When you are satisfied with the script, install Muse Code:
Prompt
sh muse-install.sh
After installation, check the version:
Prompt
muse --version
Create an empty folder called wonder-practice somewhere outside your real work projects. Open a terminal in it and start:
Prompt
muse --model muse-wonder-1.0 --reasoning-effort high --disable-shell --disable-write --sandbox-network restricted --approval-mode untrusted --approval-judge off --contributor-mode off -c context.foreign_personal_rules=false -c context.foreign_personal_skills=false
This command selects standard Wonder at High effort, disables shell execution and file writes, restricts sandbox network access, and turns off the automated approval judge for this run. It requests Contributor mode off for this run. Verify the active mode before sending a prompt. Stop if the interface still reports Contributor mode on. Confirm the account's access and data terms before generating.
It also excludes compatible personal rules and skills for this run. Before sending the brief, verify Contributor mode off in settings:
Prompt
/settings
The model ID comes from Meta’s current model listing. Confirm the selected model in the session with the model menu:
Prompt
/models
A response that calls itself Wonder is not enough evidence. Changing the effort setting also leaves the model choice intact: High or Max effort on standard Wonder keeps it standard Wonder.
Native Muse Code session export confirms Meta as the provider and muse-wonder-1.0 as the recorded model
Follow the browser sign-in prompt and check your account’s access and billing. You need not create an API key for the browser sign-in route. In Muse Code, a META_API_KEY environment variable takes priority over a stored API key, which takes priority over a stored browser session. Confirm the active billing route. Never paste a key into the conversation.
Before sending the brief, check the usage card:
Prompt
/usage
Confirm the subscription allowance you intend to use. Usage windows are last-observed data and may be absent before a first request. In that case, verify your active Code plan in Meta Accounts Center and check that no API key overrides the Code credential. Stop if an API-key warning appears or the billing route is unclear. Contributor mode off can still use a supported Code plan's allowance.
Keep the sandbox enabled and review the permission profile:
Prompt
/permissions
Select Read-only. If startup requires trusting a workspace, use only the empty practice folder after checking that it contains no project rules, hooks, plugins, or private files. The exercise needs text responses only. File reads and built-in web tools may still be available. Inspect their permissions separately; the sandbox network setting alone does not disable web search. The brief asks the model to work from pasted text, and you should review the tool log afterward.
If Wonder is unavailable, stop the Wonder-specific run and record that limit. A Spark response can be a separate comparison, but should not appear as a Wonder result.
Step 2
Give Wonder a brief with rules you can check
Paste this into the confirmed Wonder session:
Prompt
Fictional practice brief: Northline team workshop
[B1] Plan an in-person workshop for 24 attendees on October 22, 2026.
All clock times use America/New_York. The room is booked from 9:00 a.m.
to noon. Use one track, with no overlapping activities.
[B2] Include exactly these six activities, once each, without shortening
or splitting them: welcome (15 minutes), facilitator demo (45 minutes),
hands-on practice (60 minutes), break (15 minutes), group review
(30 minutes), and wrap-up (15 minutes).
[B3] Welcome comes first. The demo must finish before hands-on practice
starts. Group review starts after practice finishes. Wrap-up comes last.
[B4] The break is fixed at 10:00–10:15 a.m. The facilitator can deliver
the demo only between 9:00 and 10:00 a.m. All other activities have no
additional staff-availability restrictions.
[B5] The fixed room cost is $240 and the fixed facilitator cost is $300.
Refreshments cost $12 per attendee; materials cost $3 per attendee.
All amounts are fictional USD and include all quoted fees. Add no tax
or charge beyond the reserve in B6. The room price covers only B1's
booked time. There is no quote for extra room time.
[B6] Add a 10% reserve to the full room, facilitator, refreshments, and
materials subtotal, once. The total budget including this reserve is
$1,100. Unused reserve is still included in the planned total.
[B7] Return an agenda, itemized budget, and short list of open decisions.
Show start and end times. Cite source labels for the rules you apply.
Show the planned cost per attendee and budget remaining.
[B8] Do not change a fixed rule to make the plan fit. When a request is
impossible, explain why and ask which rule may change. Propose options
only; do not treat a suggested option as approved.
[B9] Keep everything as a draft in this conversation. Do not use tools,
read local files, browse, create files, send invitations, contact vendors,
book anything, or set a schedule. Reading this pasted brief is sufficient.
Each rule has a source label so you can trace the answer back to the brief. The same three practice prompts are included in the Pro download.
Keep the first answer before asking for revisions. Check the schedule and numbers separately. A correct budget does not fix overlapping activities.
Step 3
Check the first plan
Wonder's first native answer shows the six-activity agenda, $990 total, $41.25 per attendee, and $110 remaining
The six activities total 180 minutes, exactly the room’s three-hour booking. This agenda meets the stated order and availability:
Activity
Start
Finish
Welcome
9:00 a.m.
9:15 a.m.
Facilitator demo
9:15 a.m.
10:00 a.m.
Break
10:00 a.m.
10:15 a.m.
Hands-on practice
10:15 a.m.
11:15 a.m.
Group review
11:15 a.m.
11:45 a.m.
Wrap-up
11:45 a.m.
Noon
Check the budget against these independent calculations:
Item
Calculation
Amount
Room
Fixed
$240.00
Facilitator
Fixed
$300.00
Refreshments
24 × $12
$288.00
Materials
24 × $3
$72.00
Subtotal
$240 + $300 + $288 + $72
$900.00
Reserve
10% × $900
$90.00
Planned total
$900 + $90
$990.00
Budget remaining
$1,100 − $990
$110.00
Planned cost per attendee is $990 ÷ 24 = $41.25. The tables are independently calculated references for the fictional exercise. Compare your answer against them.
If the result differs, ask Wonder to point to the rule or calculation behind it. Repair the cause rather than asking it to copy the expected number.
Step 4
Change the headcount and check what stays fixed
Native Wonder revision keeps the agenda fixed at 30 attendees and shows a $1,089 total with $11 remaining
In the same conversation, send:
Prompt
[C1] The confirmed attendee count is now 30, replacing the 24 in B1.
Keep every other rule from the original brief, including the room hours,
activity lengths, break, facilitator availability, prices, 10% reserve,
and $1,100 budget. Revise the agenda and budget. Show what changed and
what stayed fixed. Do not take external actions.
The agenda should stay unchanged. The two per-person costs should increase:
Check
24 attendees
30 attendees
Refreshments
$288.00
$360.00
Materials
$72.00
$90.00
Subtotal
$900.00
$990.00
Reserve
$90.00
$99.00
Planned total
$990.00
$1,089.00
Budget remaining
$110.00
$11.00
Planned cost per attendee
$41.25
$36.30
Watch for a changed room fee, a forgotten materials increase, or a reserve applied only to refreshments. None follows the brief.
Save the revision before the next step. The current headcount is now 30.
Step 5
Give it a change that makes the schedule impossible
Native Wonder response explains the 12:45 p.m. earliest finish, requests approval for changes, and marks extra room costs unknown
Send this update in the same conversation:
Prompt
[C2] Keep the latest attendee count. The facilitator now says the
45-minute demo can start no earlier than 10:15 a.m. and must end by
11:00 a.m. This replaces only the facilitator-availability rule in B4.
All other constraints still apply, including the fixed break, activity
lengths, required order, and noon finish. Is a valid agenda still possible?
Explain the constraint that decides the answer. Do not shorten, overlap,
split, drop, or move an activity outside an allowed window to claim
success. Suggest the smallest changes I could approve, but do not adopt
one. Keep any extra cost that lacks a quote marked unknown.
The right answer is that the current requirements no longer fit. The demo must run from 10:15 to 11. Practice then needs 60 minutes, review 30, and wrap-up 15. The earliest finish is 12:45 p.m., which exceeds the booking by 45 minutes.
A useful response could ask whether the facilitator can restore the earlier slot or whether you can extend the room booking. It should not quietly accept either change. Extra room time also needs a quote, so it cannot claim the longer event still costs $1,089.
Check that the latest headcount stays at 30. A correct conflict explanation can still contain an old number elsewhere.
This is the part worth keeping in your own tests: a brief with no valid solution. An assistant should tell you what prevents success before inventing a way around your rules.
Step 6
Compare Wonder Full on the same task
This comparison is optional. Wonder Full requires an eligible Power or Maximum plan and regional availability. Use a fresh session, confirm muse-wonder-full-1.0 in the model menu, and repeat the original brief followed by the same two changes. Keep the first model’s answers and the reviewer key out of that session. Leave unavailable models untested.
Record model, effort, errors, retries, and time spent checking the answers. To save a conversation for review, use:
Prompt
/export transcript
Inspect a session export before sharing it; it can contain private task details.
Question
What to compare
Did it follow the initial brief?
Full activity lengths, correct order, fixed break, and budget.
Did it apply the new headcount?
Updated variable costs and unchanged fixed rules.
Did it catch the impossible request?
A correct conflict explanation without an unapproved workaround.
How much help did it need?
Corrections, retries, and your review time.
What did the full task consume?
All turns and any tool usage, under the actual billing route.
Run a separate Spark 1.3 comparison when cost matters. Keep the model that meets your checks with acceptable effort and expense. Do not change your regular workflow based on this one small example.
Keep going
Keep learning with Rundown Pro
Pro adds the full on-demand course library, certifications, group office hours, and eligible member perks.
$29/mo, cancel anytime
Is Muse Wonder free?
The confirmed routes have API usage or subscription terms. We have not verified an unrestricted free Wonder allowance. Check the model access and charges in your account before starting.
Does running Muse Code keep the task on my computer?
Muse Code is a local app that calls Meta’s hosted model service. Local installation does not make model processing local-only. Review the selected model’s data terms before adding work material.
Will the agent remember this in another conversation?
This guide tests revisions in one session. Cross-session memory needs a separate test. Muse Code has documented memory files and session controls; starting a new session does not remove saved memory or rules.
What if it starts using tools?
Interrupt the current turn and inspect what ran. In Muse Code, Esc stops the current turn but does not stop every background task or subagent. Check for work still running:
Prompt
/tasks
Keep the next attempt limited to the pasted text.
Is the API a complete assistant?
An API supplies model responses to an app. Tool access, stored context, permissions, and scheduling depend on the surrounding system. Muse Code supplies a ready-made agent environment. This exercise authorizes neither bookings nor ongoing monitoring.
