Claude Duet vs ChatGPT Dots: how two always-on agents actually differ

The clearest difference is in how each product describes its rhythm. Claude Duet sends a morning brief on the most important items across your connected apps. During the day it checks your connected accounts every few hours, and it messages you only when something needs attention. The shape of that experience is a quiet colleague who scans your inbox, calendar, and tools, and taps your shoulder rarely.

Dots are described in terms of assignments. Per OpenAI's help center as reported, a dots agent is given a goal and told what it may do independently. It asks you to decide matters that need human judgment, and you review the results it returns. The shape of that experience is closer to handing a project to a capable assistant: here is the outcome, here is how much rope you have, come back when it is done or when you must choose.

Treat this as a difference in emphasis rather than a hard boundary. Duet can plainly run tasks as well as watch, since it works on its own virtual machine, and dots retain context for ongoing work, so they can also watch a situation over time. What we can say honestly is that Anthropic's materials lead with the brief and the periodic check, and OpenAI's lead with the goal and the permissions.

Which rhythm suits you depends on your work. If your pain is the constant low-grade fear of missing something in a busy inbox, the watching model fits. If your pain is a recurring job that eats hours, such as reconciling, researching, or compiling, the delegating model may feel more natural. Many people will want both, which is a fair reason to try each when access allows.

Claude Duet is a research preview, and access is by waitlist. It is open to Claude Pro and Max subscribers in the United States, and it is not yet available on Team or Enterprise plans. If you are outside the United States, or you use Claude through a work plan, you cannot use Duet today regardless of interest.

Dots appear to have a broader opening. Per OpenAI's help center as reported, access rolled out first to ChatGPT Pro subscribers in most markets and to Business Premium users across supported regions. A beta is also available to Enterprise, Edu, and Healthcare workspaces, but an administrator must turn it on, and it is off by default. That makes dots more reachable for organizations that already live inside ChatGPT, and for individuals outside the United States.

The plan names matter more than they might seem. Duet is tied to individual Claude subscriptions, which suits a person who wants to try it for their own inbox. Dots span both individual and business tiers, which suits a person who wants to try it alone and then bring it to a team. Neither company has said how these plans will evolve, so check the current eligibility on each product's own page before you subscribe to anything for the sake of access.

One practical caution: a research preview and a beta both mean the product can change. Features, limits, and eligibility may shift between now and a general release, so avoid building a critical workflow on any detail that has not settled.

Duet's approach is concrete and action-based. 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. It also has hard limits that approval cannot override: it cannot trade securities, move money between accounts, or access someone else's accounts. On top of that, Duet ships with a dedicated classifier and safety system prompt designed to reject requests to share the strictest tier of sensitive information, such as credit card numbers and government identification.

Dots take a more permission-driven approach in the materials we could review. You set a goal and tell the agent what it may do independently, and it asks you to decide matters that need human judgment. That is flexible, since you decide how much autonomy to grant, but it also means the safety boundary depends partly on how carefully you write the instructions. We did not find a published list of hard-coded actions that dots refuse outright.

There is one related data point on the OpenAI side. OpenAI's separately documented '@ChatGPT in Slack and Microsoft Teams' feature page states that ChatGPT asks for your approval before acting in connected tools, that audit logs track agent activity, and that admin-side controls let an organization require approval for specific connected actions. That is a different integration from dots, so do not assume every control applies to dots. It does show that OpenAI treats approvals, logging, and admin controls as part of its agent story for organizations.

Put fairly: Duet publishes a clear, short list of actions that always need your approval and a few it will never take. Dots, as reported, give you a more open-ended permission model, with enterprise-style controls documented nearby. For an individual, Duet's fixed gates may feel more reassuring. For an organization that needs audit trails, the OpenAI side has more documented material today, while Duet is not yet available on Team or Enterprise plans.

Both products are new, and some of the questions you most want answered are open. No independent head-to-head test of reliability exists in the sources we checked, so we cannot tell you which one makes fewer mistakes on, say, scheduling or invoice matching. Treat any confident claim of a winner, including from vendors, as marketing until you have tried your own workflow.

Always-on also means always-connected. Whatever an agent can read, it can in principle misread, and whatever it can send, it can send to the wrong person. Approval prompts are a strong safeguard only if you read them. A habit of tapping approve without looking removes most of the protection.

There is also the question of fit with your own habits. An agent that messages you rarely and one that works through a goal both require you to describe what you want clearly. People who write precise instructions get far more out of either product. That skill, writing a clear brief with limits, outcomes, and escalation rules, transfers between Duet, dots, and whatever comes next.