Claude Duet vs OpenAI Dots vs Meta Muse: which agent fits your work?

Claude Duet, OpenAI Dots, and Meta Muse can all pursue work beyond a single chat. Start your choice with eligibility: Duet is a US Pro/Max research preview behind a waitlist; Dots requires eligible Pro or Business Premium access, or an admin-enabled Enterprise beta; Muse offers limited free usage and paid plans, with availability varying by account and region. See the Duet FAQ, Dots setup guide, and Muse subscription requirements.

Start with Duet if your approved working context is already in Claude. Evaluate Dots for ongoing work across ChatGPT and team channels. Include Muse when your task depends on its business connectors, especially Meta business accounts. These are starting points for an evaluation. Test the exact task and permissions you need before choosing; the announcements cannot establish an accuracy or reliability winner.

Verified October 22, 2026. Refix publishes this guide and offers a business-analysis and growth agent. We researched documentation, not hands-on performance. Every worked workflow below is a proposed evaluation, not a product demonstration.

What are we comparing?

Anthropic announced Claude Duet on October 21, 2026. OpenAI announced Dots on September 29, 2026. Meta introduced Muse on September 8, 2026 and added Small Business skills and connectors on September 29.

The comparison concerns Meta’s agent at muse.ai, Anthropic’s Duet preview at claude.com, and OpenAI’s Dots product. Check the company and product URL when searching: these names also appear in unrelated tools and model names.

Compare access and control before features

The linked documents support each product column. “Not specified” means the checked sources do not settle the detail. It does not mean a feature is absent.

On a phone, scroll the table horizontally to compare products.

Muse’s country descriptions have different scopes: Meta’s September Connect recap also lists Mexico for Muse generally. That does not establish that every Small Business connector or subscription is available there. Check your account before planning a rollout.

What does access cost?

US monthly list prices checked on October 22: Claude Pro is $20 and Max starts at $100, before applicable tax. A subscription alone does not clear the Duet waitlist. ChatGPT Pro tiers are $100, $200 and $500. Business Premium seats cost $125 monthly or $100 per month billed annually; Business requires two paid seats, which can mix Standard and Premium.

Muse lists Power at $20/month for 500 million tokens weekly and Maximum at $100/month for 3 billion. Benefits vary by region and account. These token allowances are not comparable units of completed work across vendors. Included access is not unlimited execution: compare the allowance consumed by your own pilot.

Duet vs Dots: how much control do you need over ongoing work?

For an existing Claude user, the practical attraction of Duet is carrying approved context into delegated tasks. Its preview also has a visible tradeoff: checks happen every few hours and you cannot yet inspect its browser. It should not be your only route for an urgent operational alert. Those are Anthropic’s stated limitations, not observations from our testing.

Dots gives you a different control surface: you can inspect its computer and manage scheduled work through its profile. Its current setup guide supports mobile-app setup, which is broader than the desktop-first launch instructions.

Choose Duet as a first evaluation if your context is already in Claude and delayed checks are acceptable. Favor a Dots evaluation if inspecting the working environment is a requirement. In either case, ask the agent to show a plan before it changes another system.

Dots vs Muse: personal delegation or a business-owned process?

Muse deserves consideration for a merchant or consumer business working across Meta accounts and a storefront. Its Small Business announcement describes campaign and financial-analysis tasks, so treating it as only a personal organizer misses relevant business work.

Evaluate Dots where the operator’s ChatGPT workspace and team channels are already approved. Its specialist dots are focused enterprise pilots. That announcement does not establish general availability for a team of departmental agents.

For either product, name the process owner before connecting apps. Confirm who receives approvals when that person is away, who can inspect the output, and what survives when an employee leaves. The checked material does not fully specify those answers for every configuration. Persistent memory alone does not establish the governance a shared business process needs.

Duet vs Muse: which connected records make the task possible?

Suppose you need an invoice follow-up. Both products name accounting or payment connections, but a connector’s presence does not prove it can retrieve the invoice fields, respect your account permissions, and draft the reminder you want.

Duet’s announcement describes finding unpaid invoices and drafting reminders. Muse documents connector authorization and optional read-only settings. Start the evaluation with one overdue invoice that you already understand. Require the agent to distinguish an open invoice from an overdue one and exclude disputed balances.

Choose by the records and permission settings your account can actually expose. Keep the first attempt in draft form. The relevant comparison is whether a reviewer can verify the recipient, amount, and evidence before anything is sent.

Design the workflow before running the agent

These are hypothetical pilot designs for a consumer business. They do not establish that each agent can access every proposed input. If a required source is unavailable, use an approved export or narrow the task.

A morning business brief

Question: What needs a decision before the team starts work?

Supply a calendar, a selected internal discussion channel, and yesterday’s reconciled orders or revenue export. Define the reporting timezone and the data cutoff. Ask for a short brief with source links, an owner for each open question, and a separate list of missing inputs.

Start with reading only and private delivery. A person checks deadlines, stale data, duplicated alerts, and whether customer details belong in the destination. An alert that arrives after a deadline fails this pilot even if its summary is correct. Duet explicitly announces a morning brief; Dots documents recurring checks; Muse describes scheduled goals. Delivery timing still needs testing in your account.

Investigating a KPI change

Question: Why did completed checkout fall while traffic stayed steady?

Provide event definitions, completed orders, payment outcomes, release notes, and comparable periods. For a marketplace, add bookable supply and cancellations. Request a reconciliation of the totals, a segment breakdown, alternative explanations, and a list of unresolved data gaps.

Keep source access read-only. Ask for an investigation memo with the calculation or query behind each material number, rather than a confident diagnosis from a dashboard screenshot. A person checks joins, refunds, timezone alignment, tracking changes, and whether the evidence establishes a cause or only a correlation. The useful output may be a better question.

Following through on a task

Question: Can the team turn an approved finding into a reviewable next step?

Give the agent the accepted finding, destination, named reviewer, deadline, and a clear stop point. For example, ask it to prepare a task describing a checkout experiment, with the affected segment, expected measurement, and rollback condition.

The proposed output is a draft task and a checklist of decisions still needed. Approval to investigate does not authorize publishing a campaign, changing prices, or sending a customer email. A person checks scope, recipients, budget, and the exact proposed change before authorizing execution. Requiring that pause is part of this pilot design; confirm that your product settings enforce it.

A team evaluation checklist

Use the same known example in each account. Record what happened rather than assigning a score from the announcements.

1. Eligibility: Can the intended operator access the product on the right plan, in their country, on their device? Duet’s FAQ explicitly excludes Team and Enterprise from the preview.
2. Source coverage: Can it read the exact records and fields needed? Check denied access as well as successful access. A marketplace total without cancellations may answer the wrong question.
3. Ownership: Name the task owner, reviewer, backup reviewer, and output destination. Confirm visibility and offboarding behavior with your administrator.
4. Repeatability: Rerun against a second period with known totals. Require a timestamp, source links, and an explanation of missing data. Record allowance use and interruptions.
5. Permission boundary: Prepare a harmless draft, then verify that the required approval appears before a consequential action. Do not test this by sending to customers or spending money.
6. Revocation and memory: Disconnect a test source and check the documented retention controls. Disconnecting access does not necessarily erase what an agent already learned.

Privacy needs a separate check

OpenAI’s Dots FAQ distinguishes Business/Enterprise/Edu no-training defaults from personal-plan settings. Disconnecting a plugin does not erase retained context; deleting a dot also leaves separately stored files and ChatGPT memories to manage.

Anthropic’s individual privacy policy permits model improvement subject to account choices, with safety-review and submitted-feedback exceptions. It explicitly separates business customer agreements. Do not apply an Enterprise guarantee to a Pro/Max Duet preview. Duet-specific retention and organization administration details are not specified in the checked launch documentation.

Meta’s security explanation describes a training opt-out. It says ordinary Secure VM operation does not prevent necessary Meta access. Confidential VM is described as forthcoming with limited testers, not a universal current guarantee. Its ad-system separation also does not mean actions taken on other websites cannot influence ads.

When should you choose each one?

Choose Duet for an initial delegated-work pilot when you have preview access, approved Claude context, and a task that can tolerate periodic checks. Use our Duet business workflow guide to specify the first task.

Choose Dots for evaluation when ChatGPT is your approved workspace and its computer visibility, scheduling, and channel options fit the work. Keep its read-only proactive research distinct from tasks you authorize to act.

Choose Muse for evaluation when the necessary business records are available through its connectors and your account’s settings support the approval boundary. Its free allowance can help with an initial bounded task, but it does not establish sufficient capacity for recurring operations.

Consider Refix when the job is investigating a business metric and pursuing a growth goal across your connected sources. Its investigations workflow tests explanations against read-only data. The Refix MCP connection lets a compatible assistant start investigations and manage watchers in one authorized cloud workspace; proposed tool actions require review in Refix. That is a specific analysis workflow. A general-purpose agent is the more natural first choice for broad inbox administration, documents, or scheduling.

Sources and limits of this comparison

Primary documents were checked on October 22, 2026. The announcement dates are historical dates, not claims about when every account received access. We have not tested these products, measured answer quality, or verified connector behavior in a signed-in account. The covers are abstract illustrations. The workflow designs and recommendations are our interpretation of the documentation.

The inline links identify the source for each material product claim: Anthropic’s announcement, product FAQ, pricing and individual privacy policy; OpenAI’s announcement, setup, safety, pricing and privacy documentation; Meta’s launch, Small Business, design, connector, subscription and security documentation. Recheck your actual access and permissions before choosing a paid pilot.