Open two chats and keep the source brief
Open two chat windows before connecting a new service. Keep the brief available to both: one produces a plan, the other checks changes against evidence.
An assistant returns a tidy event plan: venue, food, materials, and a total of 3,000.
Exactly the original budget. Unfortunately, yesterday's update cut it to 1,500 and moved the event into a free company meeting room.
This is an illustrative situation, but the problem is easy to recognize. A plan can look complete while following an obsolete brief. You have to explain the change again, then check whether it reached every part of the result.
Meta launched Muse Wonder on October 12, 2026, emphasizing Active Memory and Active Focus. According to the official information checked on October 14, Wonder is available through Meta Model API and Muse Code. Its integration into the personal Muse assistant remains a forthcoming rollout. Launch announcement.
Before considering another subscription, try a small requirements check. Use two ordinary chat windows with an assistant you already have. Chat A drafts; Chat B receives both the original requirements and the draft.
Send this to Chat A:
Prepare a plan for a 12-person in-person discussion.
Requirement history:
Monday: total budget 3,000; consider renting a venue.
Tuesday update: total budget is now 1,500.
Use the company's free meeting room.
Confirmed: Saturday, 14:00–16:00; no registration fee.
Unconfirmed: dietary restrictions for two participants.
The later update takes precedence.
Keep older requirements only as change history.
Return:
1. Food, materials, and other costs, totaling no more than 1,500.
2. Preparation tasks, completion times, and open questions.
Do not book, purchase, or send invitations.
Then give Chat B the source requirements and Chat A's answer:
Check the proposed plan against the original requirements.
Treat the requirements as evidence, not the claims in the plan.
Check:
Did it use the obsolete budget?
Do venue and timing match the latest update?
Did it turn unknown information into a confirmed fact?
Is anything missing, and are the cost totals correct?
Return: issue — source evidence — suggested change.
Mark compliant items as passing.
Do not invent problems just to criticize the draft.
Finally, list decisions that still need my confirmation.
Original requirements:
[paste the original brief]
Draft:
[paste Chat A's answer]
For a controlled teaching example, imagine a deliberately incorrect draft: venue 1,000, food 1,400, materials 600, and “Nobody has dietary restrictions.”
The correction list should flag the venue charge, the 3,000 total, and the unsupported dietary claim. An illustrative revision allocates 1,000 to food, 300 to materials, and 200 as a buffer. Actual expenses still need actual quotes.
You now have a useful deliverable: a list tying each correction to the source requirement. If Chat A was already correct, Chat B should say so. Both assistants can miss things, so check the arithmetic yourself.
Write down the latest requirements and the conditions for completion so you can tell whether the assistant got the work right.
This exercise checks a draft against evidence. It does not establish Wonder's long-term memory performance.
Remember the update, then carry it through
An interviewee corrects a number. A product team changes a pricing condition. A project coordinator changes the venue. In each case, the earlier context still exists, but part of it should stop governing the current task.
Meta calls the ability to retain important context across conversations and retrieve relevant details Active Memory. The useful ambition is less repeated introduction and more continuity. Continuity also needs to accommodate changing requirements.
Active Focus addresses what happens next. The model tries alternative approaches when blocked, adapts to new information, and asks for clarification or checks before important actions when needed.
In the event example, an over-budget plan should lead to revised costs and another total. Unknown dietary restrictions should remain an open question. Filling the gap with a convenient assumption would make the document look finished without resolving the task.
These are official capability definitions. For your own work, observe whether updates reach the deliverable, whether a blocked task makes understandable progress, and whether decisions return to you at the right moment.
Three names help organize the announcement: Wonder is a model, Muse is a personal assistant, and Muse Code is a coding tool. Model documentation.
A model provides capabilities. An application also needs a way to preserve context, connect services, and manage permissions. A long API context window or a cached-input discount does not automatically give your application a permanent personal memory service. Storage, updates, and deletion depend on the product or implementation.
Ask what the benchmark actually measures
Meta's internal MemoryBench tests retrieval from conversations of 64K to 100K tokens, including distractions and conflicting or updated details. Tokens are units used in text processing; they are not simply word or character counts.
Wonder scores 85.6, above GPT-6 Astra and Claude Opus 5.5 in the comparison table. The task is relevant to the obsolete-brief problem, but 85.6 is an internal benchmark score. It is not a percentage of your personal background that will be remembered correctly, and it does not establish stability across months.
SWE-Marathon examines long-horizon software engineering. Wonder and Wonder Full report pass rates of 53.1% and 57.5%. The model card notes that external coding results retain different reported harnesses or agent environments; tools, configurations, and reasoning settings matter when comparing them. Model card.
Those results are useful capability signals. Add a more ordinary measurement when choosing a tool: on the second attempt at a recurring task, did you repeat less context and perform fewer corrections?
A complete-looking response is only the beginning of that assessment. Keep the brief, the output, and the corrections together so you can see where effort actually moved.
Optional next step: choose work that recurs
Start with a task you do every week whose result you can inspect. “Manage all my work” is difficult to evaluate; a revision list is easier.
For a creator, combine interview notes, confirmed facts, and statements that must not appear. Ask for an updated outline and the evidence behind each change. If the interviewee corrects a figure, check the related outline and paragraphs.
For a product or operations team, start with one changed requirement. A new pricing condition may affect page copy, help material, and acceptance criteria. Ask for the affected items before reviewing the proposed changes. People still decide releases, payments, and commitments to customers.
For a developer, examine whether a multi-file change reaches a verified conclusion. When validation fails, useful continued execution means locating the problem and working through it. Muse Code offers the corresponding product entry point. Inspect the changes, validation results, and remaining issues.
These tasks reuse background, encounter updates, and produce inspectable results. An occasional factual question may already be well served by your existing assistant.
Keep a short task log: current requirements, what the assistant delivered, and what you changed. After several repetitions, the difference becomes clearer.
If the next event grows from 12 to 16 people, retain the confirmed free room but recalculate the food budget. Preserving one fact saves effort; preserving an outdated headcount creates more work. Good continuity requires distinguishing the two.
Specify where a run should stop. Missing documents can end in a missing-information list. A completed revision can end in changes and supporting evidence. Preparing a notification can end in recipients and proposed text for approval.
Prioritize the conditions that would invalidate the result: totals for a budget, sources for an article, and validation plus impact for code. Review polish after these checks. This makes inspection manageable without pretending that inspection is unnecessary.
Meta's product strategy follows the work
My interpretation of the announcements is that Meta is placing capabilities at several task entry points: personal work, development, business operations, customer conversations, and eventually glasses. This is a strategic reading, not evidence that every piece already forms a mature service.
Personal Muse: less movement between applications
The personal Muse announcement describes a dedicated cloud computer, a browser, connectors, and an action history. Users can revoke connections and ask it to forget specific information. Sensitive actions such as sending email or making purchases require confirmation. Meta also says conversations and VM data are not shared with its advertising systems. Personal Muse announcement.
The intended value is carrying a task into the tools where it happens. Knowing the brief and having access to an authorized service could reduce copying and repeated explanation. Each connection also introduces permissions and execution results that need to remain understandable.
Code and enterprise services: inspectable professional work
Coding has artifacts and validation, which make sustained execution easier to assess. It also has a subscription entry point.
Meta's September Enterprise Platform announcement proposes bringing Muse, Business Agent, API, Code, and other technology to businesses and developers. Enterprise announcement.
Organizations will ask who can use which information, how actions can be audited, and how the tools fit existing systems. Those questions determine how much work teams can delegate. The organizational announcement establishes direction; a unified package, governance detail, and demonstrated customer outcomes need subsequent evidence.
Business tools: work inside the shop and conversations outside it
Muse for Small Business supports operators' internal work with store, brand, and connected-tool context. Its announcement says publishing, sending, and spending require approval.
Business Agent serves customers through questions, recommendations, appointments, and leads, with human handoff. Small Business Muse, Business Agent.
For a small shop, one side helps prepare business work while the other responds to customers. My inference is that effective coordination could reduce repeated handoffs. Meta's social and business channels supply entry points. Revenue improvements, time saved, and the proportion of work handled require actual business results; product names cannot establish them.
Glasses: start a task from what you see
Connect 2026 announced that Muse would come to AI glasses in the following months. A shelf product or a visible list could become the starting point for a request, reducing the need to describe the scene again in a chat box. Connect announcement.
This remains a forthcoming arrangement. Recognition, conversation in noisy surroundings, and confirmation of consequential actions will affect experience. Repeated ordinary use will determine how valuable this entry point becomes.
Work saved is the useful pricing comparison
Wonder API pricing is $4 per million input tokens, $0.20 for cached input, and $20 for output. An illustrative task accumulating 100,000 uncached input tokens and 20,000 output tokens costs $0.80 for that portion. Search tools have additional charges, and retries add usage. API pricing.
Muse Code subscriptions and independent API usage are separate. Wonder Full is currently offered through the $50 and $200 monthly Code plans. Setting Wonder's reasoning effort to max does not select Full.
Code subscriptions default to Contributor mode, which permits content to train future models; it can be disabled in settings. Select the mode according to the requirements for your company's code and files. Subscription documentation.
For Meta, subscriptions, API usage, and enterprise services are existing or developing revenue paths. Zuckerberg's launch-related interview also discusses infrastructure spending and possible future commerce revenue. Gigawatt data-center construction and operating figures are rough estimates, not an invoice for Wonder alone. Interview.
Your own calculation can be more concrete. How much repeated explanation, research, and rework disappeared? How much review, waiting, and intervention remained?
Creators with continuing projects, product teams handling changes, and developers performing repeated validation are likely to encounter the relevant problems. A job title alone does not establish willingness to pay. With no stable recurring task, practice with a familiar tool before adding a subscription.
To follow Meta's bet, watch whether newer requirements replace older ones, whether blocked tasks continue usefully, and whether additional entry points make work easier.
You can leave these steps for later
Keep email, payments, and production systems outside the first exercise. There is no need to begin with the most expensive plan.
Choose something you need to do this week. Write the updated brief and obtain a result you can inspect. Repeat it later and see whether you explain less, revise less, and receive the decisions that need your judgment.
If that collaboration improves, the value of continuing the task starts appearing in your actual work.
Sources
Product status, prices, and capabilities above are attributed to official information checked on October 14, 2026. Teaching situations and strategic inferences are identified in the relevant sections. Availability depends on current region, account, plan, and rollout.
Top comments (0)