TLDR
In late 2026, three major players launched proactive “always-on” AI agents: Anthropic’s Claude Duet (October 21), Meta’s Muse (September 8), and OpenAI’s Dots (September 29). All three run on dedicated cloud computers with browsers, connect to apps, handle multi-step tasks, and seek approval for sensitive actions.
Claude Duet emphasizes professional and small-business workflows with morning briefs and strict approvals. Meta Muse targets everyday personal life (with a free tier) and WhatsApp integration. OpenAI Dots focus on persistent goals inside ChatGPT, Slack, and Teams, with customizable rules. Choice depends on your region, existing subscriptions, budget, and whether you prioritize work productivity, personal errands, or enterprise/team use. Developers building on top of these models can access Claude, GPT-series, and related models cost-effectively through unified platforms like CometAPI.
Key Takeaways
Automates morning briefings, monitors connected business applications, and prepares administrative work. However, its research preview is limited to eligible US Pro and Max subscribers with waitlist access, and users cannot directly inspect its cloud browser.
Reviewed for clarity, source attribution and current API terminology.
Claude Duet — Best for business workflows:
Meta Muse — Best for personal assistance: Helps with travel planning, shopping, scheduling, and everyday tasks, with a free entry point. Its usefulness depends on regional availability, supported service integrations, and free-tier usage limits.
OpenAI Dots — Best for long-running projects: Maintains project context, executes multi-step workflows, and lets users inspect its cloud computer. Access depends on subscription eligibility and region, while proactive background research is restricted to read-only actions.
Safety approaches differ: fixed approvals (Duet), Sentinel-style guards (Muse), and user-defined Custom Rules (Dots). None are perfect—users must still review actions.
For developers, platforms like CometAPI provide unified, lower-cost access to the underlying models (Claude Opus-series, GPT-6 Astra equivalents, and more) via a single OpenAI-compatible endpoint.
Claude Duet vs Meta Muse vs OpenAI Dots at a Glance
Traditional AI chatbots generally wait for instructions. Proactive agents extend that interaction into a continuing working relationship.
Instead of repeatedly asking an assistant to check an inbox, review a calendar, and prepare a report, a user can assign an ongoing objective. The agent then gathers information, performs permitted operations, and returns with results or questions.
Quick Comparison Table
Feature
Claude Duet
Meta Muse
OpenAI Dots
Company / Model
Anthropic / Claude Opus 5.5
Meta / Muse Spark 1.3
OpenAI / GPT-6 Astra
Launch Date
Oct 21, 2026 (research preview)
Sep 8, 2026
Sep 29, 2026
Primary Focus
Professionals & small business
Everyday personal life & goals
Persistent projects, teams, ChatGPT ecosystem
Cloud Computer
Yes (own VM + browser)
Yes (Muse Secure VM)
Yes (own Linux + Chrome; optional local access)
Interfaces
Dedicated app (iOS/Android) + claude.ai
Muse app + muse.ai + WhatsApp
ChatGPT + Slack + Teams (+ voice)
Connectors / Apps
6,000+ (Claude Marketplace)
Dozens + Meta ecosystem
4,000+ plugins
Proactive Style
Morning brief + periodic checks
Background goals + suggestions
Scheduled tasks + proactive research
Approval Model
Fixed: before send/pay/sign/delete
Sentinel gate
Custom Rules + auto-review
Access at Launch
US Pro/Max/Ultra by waitlist
US (free tier) + expansions
Pro (excl. some EU) + Business Premium + Enterprise beta
Which AI Agent Should You Choose?
Start with the task you repeat most often. For invoices, customer follow-ups, and a daily work briefing, consider Duet. For a trip itinerary or shopping research using local services, consider Muse. For a research brief or launch project that changes with new feedback, consider Dots.
Before choosing, check three things: can your account access the product, can its connected services perform the required actions, and can you review the result before an external change is made? A product that fits your workflow on paper is useful only if those conditions hold for your account.
What Is Claude Duet?
Claude Duet is Anthropic’s proactive agent, launched as a research preview on October 21, 2026. It is a standalone app (iOS, Android, and claude.ai) separate from the main Claude chat interface. Built on Claude Opus 5.5, it is designed primarily for professionals and small-business owners.
Core capabilities include:
A daily morning brief summarizing priorities across connected email, calendar, and other apps.
Background monitoring of connected accounts (every few hours to manage usage).
Messaging you only when something needs attention or a decision.
Handling tasks of varying complexity—from scheduling meetings and drafting replies to building dashboards, financial forecasts, preparing payroll, finding unpaid invoices, or prototyping web apps.
Carrying over memory, skills, and connectors from your existing Claude account.
Running on its own cloud computer/VM so work continues after you close the app.
It integrates with more than 6,000 connectors and plugins via the Claude Marketplace (examples: Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign, LinkedIn, Shopify, Canva). Sensitive actions—sending messages, making payments, signing documents, or deleting data—require explicit user approval. Some financial actions remain restricted even with approval.
Its research preview is restricted to eligible US Claude Pro and Max subscribers who receive access through the waitlist. Background checks also consume the same usage allowance as ordinary Claude activity.
Another meaningful limitation is visibility: users cannot yet directly inspect Duet's cloud browser. Anthropic also describes the current experience as stronger at professional workflows than personal errands such as booking tickets.
Meta Muse: Everyday Personal Assistance
Meta introduced Muse as a personal AI agent on September 8, 2026. Its focus includes personal goals, shopping, travel, email, and scheduling, with continued work after the user closes the app.
Instead of requiring users to understand automation software, Meta emphasizes conversational interaction through the Muse app and supported messaging channels.
Muse launched in Japan on October 23, 2026 and Singapore on October 27. Singapore partners include Shopee, Grab, Singapore Airlines, Agoda, ShopBack, Klook, Trip.com, and Lazada; connector features are rolling out progressively. Check that the particular action you need is available, not just that a service appears on a partner list.
Meta identifies Muse Spark as the agent's foundation. Its Muse Spark 1.3 release adds improvements in long-running workflows and coding, with API access through Meta Model API. That release does not, by itself, identify the exact model revision serving every Muse account.
Access and pricing: Free tier with usage limits for adults 18+ in the US (and later expansions including Canada, Japan, Singapore). Paid Power ($20/month) and Maximum ($100/month) plans increase limits.
Privacy and data-use practices have drawn scrutiny; users are typically opted into training data by default (with sanitization claims). Reliability on complex tasks varies, and Meta acknowledges mistakes will occur.
What Are OpenAI Dots?
OpenAI launched Dots on September 29, 2026, at DevDay. Dots are always-on agents powered by GPT-6 Astra. Each Dot has its own cloud computer and browser, connects to over 4,000 apps via plugins, and can pursue goals continuously.
Standout elements:
Persistent identity and memory that carries across ChatGPT, Slack, and Microsoft Teams. Dots support cloud-computer inspection and manual takeover. A connected personal computer must be online with the ChatGPT app open for local work, while cloud work can continue when that device is off. Messaging is available through ChatGPT and supported Slack and Teams connections.
Ability to open and inspect the Dot’s computer; optional connection to the user’s local machine.
Custom Rules for fine-grained control over actions (allow, require approval, block).
Auto-review layer and built-in stops for passwords, money, deletes, and installs.
Support for voice calls and planned SMS/texting.
Specialist Dots preview for organizational roles.
Dots excel at long-horizon projects, research, coding workflows (via Codex integration), and team coordination. Conversations with a Dot generally do not count against standard ChatGPT usage limits in the same way; deeper task work does.
Access and pricing: Rolling out to ChatGPT Pro plans (starting at $100/month for Pro 100, higher tiers available) outside the EEA, Switzerland, and UK, plus Business Premium seats (around $100–125 per user/month) more broadly, and Enterprise beta (admin-enabled). The first Dot is typically included.
OpenAI has been relatively transparent about limitations, including scope creep on unrelated tasks and residual risks around prompt injection or access-denied workarounds.
Agent Architecture: Execution, Memory, and Control
An agent's architecture determines where it works, how it carries context forward, which tools it can use, and when you can intervene. These choices shape everyday use: whether work continues after you close your laptop, how a project resumes, and what happens when a task needs approval.
All three agents share a foundational architecture: an isolated virtual machine in the cloud running a powerful model, a browser, storage, and connectors. The agent plans multi-step work, executes tools, maintains state, and surfaces results or requests for human input.
Claude Duet leans on judgment and prioritization. It decides urgency, drafts next steps, and stays quiet unless needed. Its strength is continuity with existing Claude context and business-oriented connectors.
Meta Muse prioritizes accessibility and personalization. The Secure VM and Sentinel create a separation between the model and external actions. Integration with WhatsApp and Meta’s social graph enables lifestyle use cases.
OpenAI Dots emphasize persistence and configurability. Custom Rules and the ability to inspect or connect local machines give power users and teams more levers. Parallel work and specialist variants point toward multi-agent collaboration.
Agent architecture
Claude Duet
Meta Muse
OpenAI Dots
Execution environment
Cloud virtual machine with its own browser; delegated work continues while your devices are off.
Dedicated Linux VM with a browser; the agent runs in an isolated container, separate from security-sensitive services.
Cloud computer with persistent files and browser sessions; optional access to a connected personal computer while it is online.
Context and persistent state
Carries over Claude account memory, skills, and connected apps; interaction continues in one ongoing thread.
The dedicated VM holds application data; durable state is stored outside the agent runtime. The cited article does not detail memory retrieval.
Uses current conversation context, relevant ChatGPT memory, and saved notes. Delegated tasks receive their own instructions and selected context.
Tool and application access
Uses connected apps, Marketplace connectors and plugins, and its cloud browser.
Uses provider-integrated connectors and skills; can build custom connectors for services with APIs or CLIs.
Uses enabled plugins within account permissions. Local skills require a connected computer; messaging connections do not grant app access.
Background work and triggers
Morning briefings and checks of connected accounts every few hours; background activity shares your plan's usage allowance.
Its VM supports background execution, scheduled jobs, and concurrent subagents.
Supports continuing assignments, saved schedules, and event monitoring where the connected service supports it.
Task orchestration
Works out steps and apps for delegated tasks. The cited announcement does not detail its internal subagent topology.
Can launch subagents and execute generated tools within its runtime; safety services remain outside that runtime.
Coordinates background agents and visible cloud threads; can create or continue supported Work and Codex tasks.
Underlying AI Models
All three isolate agent execution in cloud virtual machines so work continues independently of the user’s device.
Claude Duet runs on Claude Opus 5.5. Context (memory, skills, connectors) carries over from the main Claude account. The agent decides effort level and prioritizes what needs human attention.
Meta Muse is powered by Muse Spark 1.3 (from Meta Superintelligence Labs), trained for long-running agentic workflows with large context. Muse Secure VM isolates the agent, files, memory, and credentials; Sentinel acts as an independent gatekeeper for network actions.
OpenAI Dots use GPT-6 Astra. Architecture includes proactive research (read-only), planning/delegation to sub-agents or Codex, Custom Rules, a separate Auto-review agent, and hard stops for passwords, money movement, deletes, and installs. The cloud computer is a restricted Linux environment with Chrome.
These models represent the frontier of agentic reasoning in 2026. Developers who need to experiment with or productionize similar capabilities across providers can access Claude-family models, GPT-6 Astra, and hundreds of others through a single OpenAI-compatible endpoint such as CometAPI (https://api.cometapi.com/v1), which reports 20–40% cost savings versus direct vendor rates and sub-400 ms average latency.
Foundation-model scores can help you identify strengths in coding, tool use, and professional tasks. They do not measure how reliably Duet, Muse, or Dots completes an entire assignment. The published comparison below is useful background, not a three-agent leaderboard.
Benchmark
Claude Opus 5.5
GPT-6 Astra
Muse Spark 1.3
Terminal-Bench 4.0
66.4%
57.9%
No directly comparable result in this comparison
FrontierCode v1.1 Main
54.4%
53.3%
No directly comparable result in this comparison
AutomationBench
40.0%
41.4%
No directly comparable result in this comparison
GDPval-AA v2.1 (Elo)
1,846
1,542
No directly comparable result in this comparison
Test notes: In thepublished comparison**, Terminal-Bench uses Opus at xhigh and Astra at high effort; Astra’s score is reported by OpenAI. Other Opus results generally use max effort, with safeguards and designated fallbacks where applicable. Zapier’s AutomationBench uses no fallbacks and counts safeguard interventions as failures. GDPval-AA is an Elo rating. These are different test environments, and the summary does not fully specify hardware or runtime settings.
Opus leads the listed terminal, coding, and professional-work measures, while Astra leads AutomationBench. Muse has no directly comparable result in this table. These differences can guide what you test; they do not establish which finished agent will save you the most time.
Security Boundaries and User Control
Duet's safety and privacy protections combine model training, action review, credential handling, and checks on sensitive information. Important actions such as sending messages, making payments, signing documents, or deleting information require approval; some activities remain prohibited even with permission.
Muse's isolated runtime and credential services keep sensitive services outside the agent's execution container. Its Sentinel controls connector actions and network access. This architecture limits what a compromised agent can reach, rather than assuming every instruction it encounters is trustworthy.
Dots adds an inspectable cloud computer to its permission and approval controls. Read-only proactive research and authorized task execution have different boundaries. A large integration catalog is useful only when you understand what each connected account allows the agent to read or change.
No vendor's published security design should be interpreted as proof of zero risk.
Real-World Agent Task Evaluation
The clearest differences appear in the work each product helps you finish. Duet has published customer examples of administrative assistance; Muse emphasizes consumer tasks; Dots provides documented tools for coordinating ongoing work. The evidence below supports a choice by workflow, rather than a universal performance ranking.
Task
Published evidence
What it means for you
Evidence type
Catch business scheduling problems
Anthropic reports that Duet flagged a property showing scheduled on the wrong day and prepared a client-call brief while noticing a conflict with a flight.
Duet is a sensible first trial when missed details and daily coordination are your main burden.
Vendor-published customer accounts; no measured success rate.
Turn personal plans into actions
Meta describes Muse handling travel and email tasks, preparing grocery lists from saved recipes, and requesting approval before purchases.
Muse is a natural starting point for everyday planning, provided the services and actions you need are available locally.
Announced capabilities and examples; no comparative timing data.
Carry a changing project forward
The Dots guide documents visible work threads, follow-up instructions, and coordination of tasks as priorities change.
Dots is a strong candidate when you need to revise work in progress and inspect the outputs of several related tasks.
Documented workflow; no three-product completion benchmark.
Our assessment: prioritize Duet for administrative follow-through, Muse for everyday coordination, and Dots for evolving project work. These are recommendations based on the evidence above, not measured claims that one product is faster or more accurate. The deciding result is a usable deliverable with less effort from you.
Claude Duet vs Meta Muse vs OpenAI Dots: Cost
Subscription Prices and Usage Allowances
The prices below are in US dollars as of October 28, 2026. Subscription charges, API token fees, and the cost of a successfully completed workflow measure different things.
Claude Duet: Included at no extra charge in Claude Pro ($20/month or ~$17 annualized) and Max (from $100/month) during the research preview. Usage draws from the plan’s existing limits (5-hour sessions + weekly caps). US waitlist only; Team/Enterprise not yet supported.
Meta Muse: Free tier with usage limits; Power plan ~$20/month (higher weekly token allowance); Maximum ~$100/month. Available to adults 18+ in the US and Canada (with later expansion to additional markets).
OpenAI Dots: First Dot included with ChatGPT Pro (tiers starting at $100/month) and Business Premium (~$100–$125 per user/month annualized). Usage of certain delegated tasks may count against plan limits after an initial period. Pro availability excludes EEA, UK, and Switzerland at launch; Business Premium and Enterprise are broader.
Pricing factor
Claude Duet (Claude pricing)
Meta Muse (free and paid usage)
OpenAI Dots (ChatGPT pricing)
Individual entry route
Pro: $20/month when billed monthly
Free for many everyday needs
Pro: from $100/month
Higher capacity
Max: from $100/month
Paid plans; check local in-app offer
Pro tiers: $100, $200, $500/month
Agent access
Included in eligible Pro/Max preview access
Subject to regional rollout
Eligible Pro accounts; Business Premium and Enterprise rollout
Usage considerations
Shares Claude usage allowance
Free allowance and paid capacity
Assigned Work/Codex tasks use those products' allowances
Access limitation
US waitlist; subscription does not guarantee an invite
Service and connector availability vary
Plan, region, age, and administrator settings
CometAPI Pricing and Billing Conditions
For a custom application, start with the model and billing tier that fit the workload, then budget separately for browser execution, storage, integrations, and monitoring. The following CometAPI rates cover model input and output tokens.
CometAPI model pricing
Claude Opus 5.5 API in CometAPI
GPT-6 Astra API in CometAPI
Muse Wonder API in CometAPI
Input per 1M tokens
$3.20
$8.00
$3.20
Output per 1M tokens
$16.00
$40.00
$16.00
Rate condition
Base tier
Input length ≤272,000 tokens
Published token rates
10,000 input + 2,000 output tokens
$0.064
$0.160
$0.064
The example excludes caching, tool charges, infrastructure, and retries. Astra's longer-context tier is $16 per 1M input tokens and $60 per 1M output tokens. These are CometAPI rates checked on October 28, 2026; your effective cost depends on the request and billing conditions.
Cost per Accepted Task
Published subscription prices do not provide a reliable cost-per-task comparison.
For a subscription, divide the cost allocated to the evaluation period by the number of tasks that meet your acceptance criteria, and record human review time separately. For a custom API agent, include billed model and tool charges, retries, fallbacks, and infrastructure in that calculation. A lower monthly fee or token rate is useful only if the resulting work meets the same quality threshold.
Total cost of ownership also includes connector setup time, review overhead for approvals, and potential over-usage on background work. Muse offers the lowest barrier for casual users; Duet and Dots favor those already subscribed to higher-tier plans.
For teams building internal agents or hybrid systems that call the same underlying models, a unified gateway such as CometAPI eliminates multiple billing relationships and typically delivers 20–40% savings on model inference while remaining fully OpenAI-SDK compatible.
Which AI Agent Is Best for Different Users?
Across all three, best practice is the same: start with read-only or low-stakes permissions, review every consequential action, monitor connected accounts, and avoid granting broad financial or destructive access until reliability is proven for your specific workflows.
Small-business owners and solo professionals focused on email, invoicing, CRM, and admin → Claude Duet. Strong connector density for business tools and judgment-oriented briefing style.
Consumers and individuals wanting everyday help (travel, shopping, personal goals) with minimal cost → Meta Muse. Free tier + WhatsApp convenience.
Teams already living in ChatGPT, Slack, or Microsoft Teams who need long-running project continuity and custom guardrails → OpenAI Dots.
Developers and product teams building multi-agent or multi-model systems → Use the consumer/pro agents for end-user experiences while routing model calls through a flexible API layer (e.g., CometAPI) that surfaces Claude Opus-class, GPT-6 Astra, Muse Spark equivalents, and many specialized models under one key and dashboard. This approach supports rapid experimentation, cost control, and fallback strategies without vendor lock-in.
Geographic and plan restrictions remain binding constraints in late 2026: Duet is US waitlist-only for Pro/Max; Muse is primarily US/Canada (with expansion); Dots have broader but still gated availability.
If you are...
Start with
Why it fits
Check before committing
A small-business owner handling invoices, follow-ups, and daily priorities
Claude Duet
A work-focused briefing and administrative workflow.
Preview access, the specific accounting or CRM actions you need, and shared usage limits.
An individual planning trips, shopping, or coordinating everyday tasks
Meta Muse
Consumer-oriented tasks and a free entry point.
Local availability, supported purchase or booking actions, and free-tier capacity.
A researcher, project lead, or developer working through repeated changes
OpenAI Dots
Continuing task coordination and visible work threads.
Account eligibility, access to project sources, and the environment where each task runs.
Someone who needs to inspect and intervene in a browser workflow
OpenAI Dots
A documented inspect-and-take-over workflow.
Website compatibility and sign-in requirements; takeover does not guarantee task success.
A team needing custom business rules and repeatable integrations
A custom agent
More control over the workflow and its acceptance criteria.
Engineering, hosting, monitoring, and support costs in addition to model API charges.
Conclusion and Recommendations
Claude Duet, Meta Muse, and OpenAI Dots represent the first widely available wave of always-on personal and professional AI agents. They share a common architecture—cloud VM, browser, connectors, background execution, and human-in-the-loop approvals—while diverging in audience focus, ecosystem depth, and control granularity. No single agent dominates every use case; the best choice depends on existing subscriptions, primary workflows, and risk tolerance.
For organizations and developers who need the underlying intelligence of these systems (or want to compose custom agents), a unified access layer is increasingly practical. CometAPI aggregates 500+ models—including the families powering Duet, Dots, and related capabilities—behind a single OpenAI-compatible endpoint, offering competitive pricing, rapid model availability, and simplified operations. Visit cometapi to explore the catalog, obtain an API key, and begin building with frontier models today.
FAQs
Where Can You Access Claude Duet, Meta Muse, and OpenAI Dots?
Claude Duet uses eligible Claude Pro and Max preview access in the United States. Meta Muse is available through its supported mobile and web channels in rollout markets. OpenAI Dots is accessed through eligible ChatGPT accounts and supported messaging connections. Check your account and required integrations before committing to a recurring workflow.
For Dots, current access conditions include Pro 100, 200, and 500 users over 18 outside the European Economic Area, United Kingdom, and Switzerland. Business Premium and Enterprise are rolling out worldwide; Enterprise requires administrator enablement. An eligible account may still wait for the rollout.
What is the main difference between Claude Duet, Meta Muse, and OpenAI Dots?
Duet emphasizes professional judgment and business connectors; Muse prioritizes accessible personal assistance with a free tier and Meta/WhatsApp reach; Dots focus on persistent project work inside the ChatGPT/Slack/Teams ecosystem with highly configurable rules.
Which agent is free to try?
Meta Muse offers a free tier with usage limits. Claude Duet and OpenAI Dots require paid plans (Pro/Max or equivalent).
Can these agents make purchases or send emails without permission?
No. All three are designed to pause and request approval for sensitive actions such as sending messages, completing payments, signing documents, or deleting data.
Pricing Entry
Included in Claude Pro (~$20/mo)
Free tier; Power $20/mo
Included in Pro from ~$100/mo
Memory / Context
Carries Claude memory & skills
Persistent personal memory
Carries across platforms; learns preferences
Browser Visibility
Limited in preview
Supported
Full inspect/takeover possible
Permission and credential boundaries
Checks consequential actions before execution; credentials are kept out of the agent's direct view.
A separate Sentinel authorizes connector actions and network access; credential services sit outside the runtime.
Action review applies instructions, permissions, and safeguards. Proactive research is read-only; website sign-in uses a private flow.
Visibility and human handoff
Returns results and approval requests; the current preview does not show its cloud browser.
Clients connect to the dedicated VM. The cited technical article does not establish a browser takeover workflow.
Users can inspect the cloud computer, take over its browser, return control, and review task progress in Activity.
