Products & Services · · Team LATENT
Muse Wonder and Dots: different ways to delegate work
Separate Muse Wonder models from Muse Code and the Muse app, then compare tools, computers, memory, integrations, approvals, prices and access from Japan with Dots.
Suppose you want an AI to read email, update a briefing and keep following changes tomorrow. Model scores alone do not tell you how to delegate that work. You also need to know where code runs, how sources are connected and where approvals stop an action.
Meta announced Muse Wonder and Wonder Full on October 12, 2026 as agent-oriented models. Muse Code is a ready-made coding tool, while the Muse app takes on everyday responsibilities. Separating these offerings makes a comparison with Dots useful for choosing how to work.
Key takeaways
- Muse Wonder and Wonder Full are models with Active Memory for relevant context across conversations and Active Focus for handling obstacles, asking necessary questions, and seeking confirmation before important actions. Wonder is available through the API and Muse Code; Full is available through Muse Code Power/Maximum. Wonder’s availability in the consumer Muse app is planned.
- Muse Code provides an option for code editing and shell operations, with approvals, sandboxing, and MCP support. Computer use through the API requires an execution environment supplied by the developer. That differs from Dots, which bundles a cloud PC, integrations, and ongoing notifications.
- Wonder API pricing is US$4 for input, US$0.20 for cached input, and US$20 for output per million tokens. Muse Code enables Contributor mode by default, combining training use with increased usage allowances; users can opt out. Cost and data-use terms should suit the work being delegated.
What you obtain and what you must provide
Model → execution system → ongoing service
Model layer
Muse Wonder / Wonder Full: interpret context and produce plans and tool calls.
Execution layer
Muse Code supplies editing, shell access and approvals. An API-based agent needs your integrations and runtime.
Ongoing-service layer
Muse app / Dots: delegate work with memory, connections and updates. Wonder in the Muse app is forthcoming.
Meta October 12 announcement and planned Muse-app rollout
Separate the model from the place you delegate work
Wonder is available through Meta Model API and Muse Code. Wonder Full is offered in Muse Code Power Usage and Maximum Usage, not through the API. Full is a separate model, not ordinary Wonder with its reasoning effort set to max. Wonder’s rollout in the consumer Muse app is planned to begin with subscribers.
Developer announcement: access to Wonder and Full
Model catalog: Full versus reasoning settings
Dots is an ongoing-agent product powered by GPT-6 Astra with a cloud computer and browser. Users assign responsibilities, connect sources and receive results or requests for decisions. With a custom Wonder API agent, developers design the tools exposed to the model and the system that executes them.
Dots product components and ongoing work
What Active Memory and Active Focus address
Active Memory is described as retaining relevant context across conversations and retrieving it for the current task. Active Focus concerns trying alternatives and adapting to obstacles or new information during difficult work, including necessary questions and checks before important actions.
Updating a meeting brief may require both recalling an agreed deadline and finding another source when a document is inaccessible. The two capabilities address those different needs. Storage, retention, wake-up timing and app permissions still depend on the product or implementation.
Definitions of Active Memory and Active Focus
The model card lists a one-million-token context, text/image/video/audio inputs and text output. Availability of each input, including audio, must be checked on the actual API or app surface. A large context window is also distinct from a product’s persistent-memory management.
MemoryBench and FocusBench include internal Meta evaluations, and SWE-Marathon compares results from different agent harnesses. Scores do not establish that the Muse app remembers more reliably or operates autonomously longer than Dots. Deliverable quality and the effort spent supervising the intended workflow need separate evaluation.
Model card: input/output and evaluation conditions
Compare tools, computer use and app connections
| Offering | Tools and computer use | External app connections |
|---|---|---|
| Wonder API | Tool calling; for computer use, the developer supplies the target environment, driver and screenshots. | Build function tools and integrations. Model access alone does not connect every service. |
| Muse Code | Edits and shell commands in a local project; macOS, Linux and Windows. | Extend with MCP, skills, hooks and subagents, with authentication and permissions for each service. |
| Muse app | Designed around a Muse Secure VM for browser, file and code work. | Connectors include Slack, Shopify and accounting services; verify each connection’s scope. |
| Dots | Own cloud computer and browser; can also delegate work to an optionally connected local computer. | Uses supported plugins and connected accounts. A messaging channel and app-action permissions are separate. |
Execution infrastructure required for API computer use
Muse Code editing, shell and OS support
Muse app Secure VM and permission controls
Muse app connectors and availability
Dots cloud computer and local connections
“Computer use” covers different arrangements. Meta’s computer-use API does not bundle a hosted VM or browser. Developers supply the runtime, action driver, checks, logs and recovery. Muse Code is a ready-made CLI, so its setup burden differs from building that API loop from scratch.
Long-running work and approvals
Muse Code provides project instructions and persistent local memory for work across sessions, with approvals and an OS sandbox enabled from first use. Windows uses PowerShell; voice input and inter-session messaging are explicitly unavailable there. Workflow availability varies by build and rollout.
Muse Code configuration and memory
Muse Code approvals and platform differences
Dots tracks work between conversations and can resume as needed. Its cloud work can continue while the user’s computer is off. Local work requires the connected computer to be online with the ChatGPT app open. Continuing in the cloud does not mean controlling local apps while that computer is off.
Dots ongoing tasks and retained context
Memory also differs at the product level. Dots uses relevant ChatGPT memory and its own saved notes, not a complete transcript of everything said. The Muse app maintains memory across conversations and lets users inspect and edit memory files. Neither mechanism can be described by the model’s input limit alone.
Muse app ongoing work and editable memory
Example: update a customer brief
Separate execution from review
Research and prepare
Read authorized sources → prepare changes → preserve sources and edits.
Return a decision
Apply approval controls when recipients, publication or purchases are involved.
Define continuation
Check when work resumes, what triggers updates and how loss of budget or access stops it.
The Muse app documents approval before publishing, sending or spending. Dots combines existing app permissions with automatic action review to decide whether to proceed, request approval or hand a step to the user. A model’s stated tendency to check before important actions does not mean every product operation always waits for manual approval.
Dots action review and stopping controls
Stopping behavior also matters. Dots Pause stops its main task; delegated tasks and recurring schedules have separate controls, and completed actions are not undone. A trial should check whether any sends or updates remain pending after a stop.
Read pricing and data-use conditions
| What is purchased | Verified pricing and allowances |
|---|---|
| Wonder API | Per million tokens: $4 input, $0.20 cached input, $20 output. Web search adds $2.50 per 1,000 queries. No long-context premium. |
| Muse Code | Monthly Everyday / High / Power / Maximum plans, with 1 / 5 / 20 / 80 times baseline usage. Currency-denominated monthly prices were not confirmed in the inspected documentation. |
| Muse app | Free use and additional subscriptions. Treat these separately from Muse Code plans and API token prices. |
| Dots | Check the eligible ChatGPT plan and allowances. Conversation usage and delegated Work / Codex usage are counted separately. A standalone Dots price was not established here. |
Meta Model API pricing and additional charges
Muse Code subscription allowances and Contributor mode
Dots eligible plans and usage allowances
Muse Code subscriptions default to Contributor mode. Increased allowances come with permission to train future Meta models on submitted prompts, code, files, context and responses. Turning it off prevents future content from being used for training. Check this setting before submitting business code.
Wonder API has no Contributor mode and uses Standard pricing. Muse Code subscription credentials are also distinct from separately issued pay-as-you-go API keys. An 80-times usage allowance does not guarantee 80 times the runtime on a task; workload and retries affect consumption.
Check what is available from Japan
The official Muse business page lists availability for adults 18 and over in the US and Canada. Japanese general availability is not established. API and Muse Code access have separate conditions. Subscription benefits vary by region; the inspected public documentation did not settle Japanese eligibility or yen pricing.
Muse app regional and age requirements
Muse Code regional subscription conditions
Dots is gradually rolling out to Pro 100, 200 and 500 users over 18 outside the EEA, UK and Switzerland; Japan is not in that exclusion. Business Premium and Enterprise are rolling out worldwide, with administrator enablement required for Enterprise. An eligible plan does not guarantee immediate access.
Current Dots access conditions
The October 9 update also documents creating a dot in the ChatGPT mobile app. That establishes a mobile-app feature, not equivalent mobile-browser functionality. Japanese-language task quality should be assessed separately from regional access.
October 9 update: mobile creation and Codex integration
Choose where to delegate each kind of work
For repository changes, Muse Code supplies editing, execution, approvals and memory. For an internal system with custom integrations and logs, Wonder API is a candidate. For ongoing work that tracks customer feedback, updates materials and requests decisions, compare services such as Dots and the Muse app.
This extends our earlier distinction between producing a deliverable and owning an ongoing responsibility. Choosing a new model does not settle notifications, connections and spending controls. Compare one bounded workflow on deliverable accuracy, review effort, recalled context and the state after stopping.
Related: Gemini agent, Dots, Muse and other work agents
Related: ongoing responsibilities with Dots
Coding work can be divided further. Our ColGrep-Agent article examines delegating only search to a small local model. Separating model, tools and continuity also makes it possible to consider arrangements beyond assigning everything to one product.
Related: delegate code search to a local 2B model
Sources and remaining unknowns
Comparison date: October 13, 2026. This is a review of official announcements, the model card and product/API documentation, not a measured identical-task comparison. Unconfirmed means not established by the inspected sources, not unsupported. Internal tool specifications and private user environments are not used as product evidence.
Japanese Wonder API and Muse Code eligibility, actual Muse Code monthly prices, per-account Wonder rollout dates and long-running Japanese task quality remain unconfirmed. Check regional eligibility and data settings at signup, then begin with a bounded reading-and-drafting trial.