Claude Duet vs OpenAI Dots: Which Agent Asks First?

People typing claude dots into a search box are not looking for a product that exists. They have heard of OpenAI Dots, the persistent assistant OpenAI launched on 29 September 2026, and they are checking whether Anthropic shipped the same thing under a similar name. It has not, and the two are unrelated: Claude Duet is Anthropic's own agent app, announced on 21 October 2026 and designed around Claude Opus 5.5. Different companies, different products, different everything except a syllable.

Dots is a named assistant that lives inside ChatGPT and gets its own cloud computer. Duet is a separate app you install. Neither company has licensed anything to the other, and neither name is a feature of the other product. A page that calls Claude Duet Anthropic's answer to Dots is describing a rivalry, not a lineage. What follows compares the two on the dimension that actually separates them — what each agent does before it touches something that cannot be undone. Everything about Dots comes from Vellum's reading of OpenAI's launch documentation, published 29 September 2026, and is attributed that way every time it appears. Everything about Duet comes from Anthropic's own announcement of 21 October 2026 and is stated directly.

Two products, two defaults

Per Vellum's reading of OpenAI's docs, Dots arrived as a named, persistent assistant inside ChatGPT, running on GPT-6 Astra, with a hosted computer and browser so that work continues when your laptop is shut. The same reading puts its reach at more than 4,000 apps, with context carried between ChatGPT, Slack and Microsoft Teams and the work inspectable as it runs. One primary dot is the starting configuration; a team of specialist dots is an enterprise pilot rather than a standard feature.

Claude Duet is a standalone app on iOS, Android and claude.ai, opened as a research preview by waitlist on 21 October 2026 for Pro and Max subscribers in the United States. It inherits your Claude memory, skills and connected apps, works in a virtual machine with its own browser, and messages you when a task completes or needs review. Anthropic's own limitation list is short and specific: it may misread a thread, personal errands are weaker than work tasks, it cannot yet show you its browser, and it lives in its own app rather than inside the main Claude app.

The difference that matters: the step before the irreversible one

Both products are agents: both can look at your accounts and act on what they find. The distinction worth paying for is not capability but where each vendor drew the line between looking and acting.

• OpenAI Dots, per Vellum's reading of OpenAI's docs — proactive research is read-only. While the dot is hunting for useful work in connected apps it cannot send messages, change app content, or control a browser or computer. User-directed and scheduled tasks can act, but under separate rules and approvals.

• Claude Duet, per Anthropic's announcement — it reads, drafts and prepares work on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits for your approval.

Those are not the same rule, and it would be dishonest to call either one stricter without saying where. Dots' read-only boundary covers the unsupervised sweep; the moment you request a task, or a schedule fires, approvals come back into play. Duet draws its boundary around the action class instead: read and draft freely, stop before anything lands. For someone handing over a shared inbox, the practical question is which shape fits the work. A dot that never acts unless asked is safer in one sense and less useful in another. An agent that drafts all day and halts at the send button is closer to a colleague with a review step.

What is refused even with your approval

An approval gate describes what an agent asks about. A quieter design question is what it declines no matter what you tell it. Anthropic answered that one on the page: a few things are off-limits even with your approval, including trading securities and moving money between accounts. Those are not prompts you can click through. They are out of scope.

OpenAI's equivalent, per Vellum's reading of the docs, sits on the user side rather than the product side — some sensitive tasks, including changing a password, remain with the user. Vellum also records automatic review for actions that affect accounts or share information, custom rules a user defines, and monitoring that can pause suspicious work. That is a real set of controls, documented in more detail than most vendors manage at launch, and it deserves credit. What neither vendor has published is a failure rate for any of it. Vellum says so plainly about Dots — the design controls are meaningful, their real-world failure rates have not yet been independently established — and the same sentence applies to Duet's approval gate, which shipped this week.

Credentials, and the classifier Anthropic built for the awkward cases

Duet's handling of credentials follows the design Anthropic uses for Claude Managed Agents, so the agent never directly sees your passwords or sign-in tokens. On top of that, Anthropic released a dedicated classifier and safety system prompt that checks each message for sensitive-information leaks. Its job is to reject requests to share the strictest tier of information, including credit card numbers and government identification; health and personal finances get through only when you have explicitly instructed the agent to send them. Red teams and an external firm penetration tested the product, and it is covered by Anthropic's public bug bounty program.

Dots reaches a similar end by a different route. Per Vellum's reading, supported saved-password flows keep credentials out of the model's context, so the model is not reading your password either — but the mechanism is a credential flow, not a classifier inspecting outbound content. Which of those holds up better against a persuasive prompt injection is exactly the question a launch-week page cannot settle. Anthropic at least named the threat it was designing against, citing its browser-use research on hidden instructions. Vellum's analysis of Dots names injection risk only through the auto-review and monitoring controls, which respond to suspicious behaviour rather than screening what the model is about to emit.

Retention: the one place Dots is better documented

Here the fair thing to write is that OpenAI's published rule is sharper than Anthropic's. Per Vellum's reading of OpenAI's docs, disconnecting an app does not erase information the dot already learned from it, and resetting the dot deletes its conversations, memories and scheduled tasks in one operation. That is an uncomfortable rule. It is also a rule, stated plainly, with two different operations defined and their different consequences spelled out. A buyer can evaluate it, argue with it, or decide it is disqualifying.

Anthropic says you are able to disconnect any app Claude Duet has access to. It does not say what happens to what Duet already learned from that app, and it describes no equivalent reset operation covering memory and scheduled tasks. That is a gap in the documentation, not evidence of worse behaviour; the honest reading is that Duet's retention semantics are unpublished. If deletion rules matter to your business — and for anything touching client records they should — that is a question to put to Anthropic before you connect the account, and a different one to put to OpenAI before you connect a different one.

Reach, in the first weeks

Neither product contacts you the way a colleague would. Per Vellum's reading, you can call your dot but it cannot initiate a call to you at launch, and texting is a limited US Pro beta rather than a general channel. Duet messages you — a notification when a task completes or needs your review — but it lives in its own app, separate from the main Claude app, and access is a US Pro and Max waitlist.

The regional shape decides whether any of this applies to you. Per Vellum's reading, Dots' Pro availability excludes the EEA, the UK and Switzerland at launch, with an admin-enabled beta for Enterprise, Edu and Healthcare workspaces. Duet starts with Pro and Max users in the United States. Both are early and narrower than the announcements make them sound.

Both ecosystem numbers are vendor numbers

Anthropic says Claude Duet can use the more than 6,000 connectors and plugins in the Claude Marketplace. OpenAI says its plugins connect to more than 4,000 apps. Both are vendor statements about their own catalogues, and neither is verifiable from outside: a listed connector is not one that works against a specific account's permissions. Treat 6,000 against 4,000 as evidence that both ecosystems are large, not as a score. The named examples are more useful than the totals. Duet's page works through Gmail and Slack triage, QuickBooks and Stripe invoices, Gusto payroll, Docusign agreements, HubSpot follow-ups, Canva campaigns and Shopify store setup, and describes checkout through Stripe's Link wallet and Shop Pay, where the agent confirms the item and the wallet separately asks you to approve the spend.

The cost question neither vendor answered

Both products are bundled rather than metered, and neither vendor has published what a completed task costs. That gap, not any feature difference, is the headline for anyone budgeting.

On OpenAI's side, the first dot is included at no extra charge with eligible ChatGPT Pro and Business Premium plans, and OpenAI has not published the ongoing Dots allowance or the price of additional dots. There is a related change that is easy to mistake for an answer: on 29 September 2026, OpenAI's Tibo Sottiaux said the $200 Pro subscription would reopen to new subscribers with a new usage calculation representing roughly half the API-dollar equivalent of the old plan, and that the five-hour usage limit would not return. Per Vellum's reading, that is a change to the wider Pro economics, not a disclosed post-introductory allowance for Dots.

On Anthropic's side, Duet is included in Pro and Max during the research preview and shares your plan's usage limits with the rest of Claude. It spends them differently: background checks count toward those limits, and larger tasks such as building a financial model or a deck require more model turns and more background work. Anthropic explains the every-few-hours cadence as a cost control — checking nonstop would use up the plan's limits much faster. That is a useful admission, and it is not a price.

So the comparison on cost is a draw between two unknowns, and the advice is identical: run one real task for a week and watch what it does to your plan before committing a team. Neither vendor has given you a number to check that against.

If you would rather build the loop yourself

Neither product can be pointed at a different model provider. Dots runs on OpenAI's model stack inside ChatGPT; Duet is designed around Claude Opus 5.5 and lives in Anthropic's app. If you want the agent loop and you also want to choose what runs underneath it — or you want the same model available on the day your agent vendor's plan limits stop you — that is a different purchase.

Claude Opus 5.5 is available on OrcaRouter at $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context window and cache reads at $0.20 per million, through one OpenAI-compatible key that reaches more than 200 models, with no markup over provider list price and automatic failover across providers. That is the model as an endpoint rather than as an app. To be precise about what this is not: OrcaRouter does not host Claude Duet or OpenAI Dots, and you cannot run either product through our API. What you can do is build the loop around the same model yourself, and keep it running when a consumer plan would have stopped it.

What to check before you connect an account

Take the product you are leaning toward and ask it the same four things. Find the sentence that says what it may do without asking, and check whether that sentence covers scheduled work as well as research. Find the list of refusals, and see whether it is a list of things the product won't do or a list of things you can override. Find the deletion rule, and notice whether disconnecting an app removes what was already learned. Find out whether background checks spend your plan's limits, and at what rate.

Both vendors have answered some of those questions and left others open. Anthropic is unusually forthcoming about the safety architecture and unusually quiet about retention and price per task. OpenAI is unusually clear about the two deletion operations and unusually quiet about the ongoing allowance. Neither silence is sinister; both are the kind of thing that gets published after a few months of customer questions rather than on day one.

What you can say now, without hedging, is that the two products are not the same product and the naming collision is not a rebrand. One is a managed assistant inside ChatGPT that researches read-only and acts under separate rules. The other is a separate app that drafts continuously and stops at the last step before anything leaves your account. If the confusion that brought you here was whether Anthropic cloned Dots, the answer is no, and the more interesting question is that the two companies chose different moments to interrupt the agent — which tells you something about what each one expects to go wrong.